{
  "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": 0, "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": "A9B6BC82A7AE4944ACE63267A622F373",
  "traceEvents": [
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: DivBackward0", "pid": 1336753, "tid": 1381210,
    "ts": 1514656708309.464, "dur": 128.419,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656708346.082, "dur": 83.772,
    "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": 1336753, "tid": 1381210, "ts": 1514656708346.082,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::div", "pid": 1336753, "tid": 1381210,
    "ts": 1514656708353.594, "dur": 74.423,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656708451.383, "dur": 242.252,
    "args": {
      "External id": 2483204,"Record function id": 0, "Ev Idx": 3
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward", "pid": 1336753, "tid": 1381210,
    "ts": 1514656708502.171, "dur": 104.877,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656708548.435, "dur": 46.977,
    "args": {
      "External id": 2483206,"Record function id": 0, "Ev Idx": 5
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AddBackward0", "pid": 1336753, "tid": 1381210,
    "ts": 1514656708612.381, "dur": 1.526,
    "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": 1336753, "tid": 1381210, "ts": 1514656708612.381,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336753, "tid": 1381210,
    "ts": 1514656708618.001, "dur": 70.995,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656708636.978, "dur": 51.313,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656708648.744, "dur": 3.771,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656708703.508, "dur": 36334.978,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656708705.095, "dur": 36323.390,
    "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": 1336753, "tid": 1381210, "ts": 1514656708705.095,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514656708743.752, "dur": 4.169,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656708753.090, "dur": 36139.404,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656708754.971, "dur": 36137.085,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656708759.635, "dur": 7.102,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656708768.397, "dur": 36122.508,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656744897.058, "dur": 0.514,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656744899.972, "dur": 2.697,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656744901.385, "dur": 1.106,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656744909.037, "dur": 32.104,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656744949.732, "dur": 68.456,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656744951.402, "dur": 66.565,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656744952.814, "dur": 64.533,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745050.794, "dur": 18.842,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745052.469, "dur": 14.113,
    "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": 1336753, "tid": 1381210, "ts": 1514656745052.469,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656745056.295, "dur": 10.067,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745060.356, "dur": 5.797,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745075.460, "dur": 94.621,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745076.689, "dur": 86.759,
    "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": 1336753, "tid": 1381210, "ts": 1514656745076.689,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 1336753, "tid": 1381210,
    "ts": 1514656745080.112, "dur": 82.956,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745084.569, "dur": 32.658,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745088.946, "dur": 5.184,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745095.518, "dur": 21.402,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745098.200, "dur": 18.340,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745121.524, "dur": 4.271,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745123.993, "dur": 1.415,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745127.009, "dur": 35.243,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745174.693, "dur": 61.948,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745175.571, "dur": 58.132,
    "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": 1336753, "tid": 1381210, "ts": 1514656745175.571,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select_backward", "pid": 1336753, "tid": 1381210,
    "ts": 1514656745178.426, "dur": 55.037,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745183.411, "dur": 20.886,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745184.848, "dur": 2.786,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745188.548, "dur": 15.456,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745191.813, "dur": 11.784,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745208.000, "dur": 5.126,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745211.599, "dur": 0.852,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745213.858, "dur": 19.071,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745240.764, "dur": 131.447,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745241.618, "dur": 126.994,
    "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": 1336753, "tid": 1381210, "ts": 1514656745241.618,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 1336753, "tid": 1381210,
    "ts": 1514656745245.561, "dur": 122.658,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745249.010, "dur": 17.347,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745250.064, "dur": 2.442,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745253.175, "dur": 12.915,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745253.836, "dur": 11.895,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745270.201, "dur": 1.941,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745271.270, "dur": 0.718,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745272.814, "dur": 94.560,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745376.285, "dur": 102.804,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745377.296, "dur": 99.244,
    "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": 1336753, "tid": 1381210, "ts": 1514656745377.296,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 1336753, "tid": 1381210,
    "ts": 1514656745378.536, "dur": 97.610,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745381.725, "dur": 18.733,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745382.744, "dur": 4.431,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745388.100, "dur": 12.124,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745388.864, "dur": 10.978,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745401.807, "dur": 4.584,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745405.771, "dur": 0.510,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745410.048, "dur": 65.317,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745483.004, "dur": 37.742,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745484.093, "dur": 1.094,
    "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": 1336753, "tid": 1381210, "ts": 1514656745484.093,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336753, "tid": 1381210,
    "ts": 1514656745487.674, "dur": 30.312,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745489.583, "dur": 27.953,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745496.699, "dur": 2.732,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745525.624, "dur": 2269.197,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745538.966, "dur": 2225.108,
    "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": 1336753, "tid": 1381210, "ts": 1514656745538.966,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514656745576.617, "dur": 3.218,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745582.885, "dur": 2086.659,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745584.941, "dur": 2084.278,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745587.550, "dur": 6.144,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656745594.722, "dur": 2073.610,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747672.971, "dur": 0.413,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747678.399, "dur": 2.631,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747679.715, "dur": 1.132,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747684.979, "dur": 24.129,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747714.816, "dur": 41.613,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747715.865, "dur": 40.321,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747717.063, "dur": 38.767,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747772.864, "dur": 18.548,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747802.910, "dur": 15.570,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747804.285, "dur": 12.110,
    "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": 1336753, "tid": 1381210, "ts": 1514656747804.285,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656747811.134, "dur": 5.055,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747812.458, "dur": 3.567,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747822.086, "dur": 63.882,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747823.191, "dur": 58.189,
    "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": 1336753, "tid": 1381210, "ts": 1514656747823.191,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 1336753, "tid": 1381210,
    "ts": 1514656747824.802, "dur": 56.258,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747828.749, "dur": 21.406,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747830.054, "dur": 2.922,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747833.866, "dur": 16.041,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747835.727, "dur": 13.809,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747851.353, "dur": 5.243,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747853.051, "dur": 3.331,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747858.050, "dur": 22.287,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747891.920, "dur": 53.512,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747893.059, "dur": 49.874,
    "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": 1336753, "tid": 1381210, "ts": 1514656747893.059,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select_backward", "pid": 1336753, "tid": 1381210,
    "ts": 1514656747896.975, "dur": 45.729,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747898.685, "dur": 19.942,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747899.514, "dur": 2.274,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747902.481, "dur": 15.905,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747903.406, "dur": 14.547,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747919.720, "dur": 6.245,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747924.664, "dur": 0.809,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747926.875, "dur": 15.304,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747951.464, "dur": 127.006,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747955.344, "dur": 119.082,
    "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": 1336753, "tid": 1381210, "ts": 1514656747955.344,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 1336753, "tid": 1381210,
    "ts": 1514656747956.924, "dur": 117.127,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747957.824, "dur": 17.104,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747958.608, "dur": 2.199,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747961.662, "dur": 13.008,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747962.906, "dur": 11.324,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747976.381, "dur": 31.038,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656747980.388, "dur": 26.109,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656748008.977, "dur": 64.254,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656748084.417, "dur": 97.085,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656748088.603, "dur": 73.832,
    "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": 1336753, "tid": 1381210, "ts": 1514656748088.603,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 1336753, "tid": 1381210,
    "ts": 1514656748092.478, "dur": 69.718,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656748094.040, "dur": 17.383,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656748094.776, "dur": 2.231,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656748097.880, "dur": 13.278,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656748099.476, "dur": 11.178,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656748112.243, "dur": 1.821,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656748113.218, "dur": 0.720,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656748115.005, "dur": 46.604,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656748166.266, "dur": 13.927,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656748185.628, "dur": 39.131,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656748189.774, "dur": 0.749,
    "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": 1336753, "tid": 1381210, "ts": 1514656748189.774,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336753, "tid": 1381210,
    "ts": 1514656748193.162, "dur": 29.230,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656748194.993, "dur": 26.922,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656748201.719, "dur": 2.553,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656748229.349, "dur": 3071.992,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656748234.171, "dur": 3040.756,
    "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": 1336753, "tid": 1381210, "ts": 1514656748234.171,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514656748265.211, "dur": 2.378,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656748269.829, "dur": 2913.989,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656748271.681, "dur": 2911.900,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656748274.302, "dur": 3.377,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656748278.324, "dur": 2904.390,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751187.200, "dur": 0.271,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751188.603, "dur": 2.448,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751189.801, "dur": 1.102,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751194.734, "dur": 20.622,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751225.030, "dur": 43.097,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751225.997, "dur": 41.945,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751227.142, "dur": 40.494,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751282.333, "dur": 15.939,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751308.180, "dur": 10.025,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751309.428, "dur": 6.787,
    "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": 1336753, "tid": 1381210, "ts": 1514656751309.428,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656751311.869, "dur": 4.147,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751312.741, "dur": 3.093,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751321.612, "dur": 65.251,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751322.587, "dur": 60.483,
    "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": 1336753, "tid": 1381210, "ts": 1514656751322.587,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 1336753, "tid": 1381210,
    "ts": 1514656751326.366, "dur": 56.345,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751331.600, "dur": 19.952,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751332.760, "dur": 2.348,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751335.886, "dur": 15.381,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751337.013, "dur": 13.843,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751353.169, "dur": 5.416,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751357.551, "dur": 0.813,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751359.597, "dur": 22.512,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751390.736, "dur": 55.903,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751391.649, "dur": 52.505,
    "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": 1336753, "tid": 1381210, "ts": 1514656751391.649,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select_backward", "pid": 1336753, "tid": 1381210,
    "ts": 1514656751393.731, "dur": 50.204,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751395.503, "dur": 22.370,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751398.743, "dur": 2.119,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751401.388, "dur": 16.255,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751402.639, "dur": 14.657,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751419.040, "dur": 8.808,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751424.671, "dur": 2.722,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751428.509, "dur": 14.879,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751450.220, "dur": 95.700,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751451.300, "dur": 90.889,
    "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": 1336753, "tid": 1381210, "ts": 1514656751451.300,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 1336753, "tid": 1381210,
    "ts": 1514656751452.682, "dur": 89.050,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751453.640, "dur": 18.366,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751454.739, "dur": 2.147,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751459.720, "dur": 12.004,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751460.788, "dur": 10.475,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751472.823, "dur": 4.204,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751476.121, "dur": 0.696,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751478.095, "dur": 62.561,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751551.831, "dur": 109.304,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751553.149, "dur": 90.650,
    "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": 1336753, "tid": 1381210, "ts": 1514656751553.149,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 1336753, "tid": 1381210,
    "ts": 1514656751557.512, "dur": 86.043,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751558.915, "dur": 23.577,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751560.050, "dur": 2.472,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751563.123, "dur": 19.117,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751566.190, "dur": 15.706,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751583.590, "dur": 4.612,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751587.356, "dur": 0.563,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751588.866, "dur": 54.119,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751646.997, "dur": 13.006,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751667.280, "dur": 30.994,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751668.616, "dur": 0.988,
    "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": 1336753, "tid": 1381210, "ts": 1514656751668.616,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336753, "tid": 1381210,
    "ts": 1514656751671.000, "dur": 23.154,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751672.607, "dur": 21.077,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751676.604, "dur": 0.445,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751702.297, "dur": 3096.106,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751712.553, "dur": 3058.145,
    "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": 1336753, "tid": 1381210, "ts": 1514656751712.553,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514656751739.310, "dur": 1.982,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751743.275, "dur": 2945.453,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751744.786, "dur": 2943.662,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751748.880, "dur": 3.022,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656751752.940, "dur": 2934.498,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754692.085, "dur": 0.278,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754693.606, "dur": 2.317,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754694.614, "dur": 1.182,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754699.037, "dur": 22.843,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754725.947, "dur": 38.541,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754726.899, "dur": 37.412,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754728.307, "dur": 35.715,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754780.272, "dur": 13.775,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754808.848, "dur": 11.320,
    "args": {
      "External id": 2483419,"Record function id": 0, "Ev Idx": 218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656754811.024, "dur": 7.505,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754814.144, "dur": 3.656,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754814.855, "dur": 2.826,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754823.408, "dur": 10.280,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754824.264, "dur": 7.427,
    "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": 1336753, "tid": 1381210, "ts": 1514656754824.264,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656754826.486, "dur": 5.009,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754828.991, "dur": 2.327,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754837.179, "dur": 61.377,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754837.954, "dur": 56.517,
    "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": 1336753, "tid": 1381210, "ts": 1514656754837.954,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 1336753, "tid": 1381210,
    "ts": 1514656754839.193, "dur": 54.961,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754844.020, "dur": 19.749,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754845.227, "dur": 2.553,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754848.407, "dur": 15.095,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754849.339, "dur": 13.689,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754865.104, "dur": 5.347,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754869.318, "dur": 0.901,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754871.249, "dur": 22.250,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754902.414, "dur": 55.369,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754903.409, "dur": 52.066,
    "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": 1336753, "tid": 1381210, "ts": 1514656754903.409,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select_backward", "pid": 1336753, "tid": 1381210,
    "ts": 1514656754904.922, "dur": 50.272,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754906.417, "dur": 26.708,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754914.117, "dur": 2.227,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754916.907, "dur": 15.975,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754917.595, "dur": 14.924,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754934.010, "dur": 4.397,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754937.009, "dur": 0.918,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754938.998, "dur": 15.771,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754961.233, "dur": 123.513,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754962.288, "dur": 119.034,
    "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": 1336753, "tid": 1381210, "ts": 1514656754962.288,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 1336753, "tid": 1381210,
    "ts": 1514656754963.814, "dur": 117.228,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754967.506, "dur": 43.908,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754968.167, "dur": 1.645,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754972.588, "dur": 38.488,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656754973.355, "dur": 37.009,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755012.951, "dur": 4.974,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755014.352, "dur": 3.313,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755018.838, "dur": 61.179,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755090.440, "dur": 115.849,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755091.313, "dur": 96.525,
    "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": 1336753, "tid": 1381210, "ts": 1514656755091.313,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 1336753, "tid": 1381210,
    "ts": 1514656755092.468, "dur": 95.129,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755093.735, "dur": 42.350,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755099.015, "dur": 2.347,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755117.178, "dur": 18.641,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755122.794, "dur": 12.667,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755137.073, "dur": 2.049,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755138.273, "dur": 0.672,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755139.647, "dur": 47.295,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755191.591, "dur": 13.025,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755212.028, "dur": 355.727,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755213.216, "dur": 345.400,
    "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": 1336753, "tid": 1381210, "ts": 1514656755213.216,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 1336753, "tid": 1381210,
    "ts": 1514656755377.485, "dur": 39.894,
    "args": {
      "External id": 2483470,"kernel_hash": "c2likoimxa3yi6j6dwntng3kzuvmiu64h4ync66w4mrx4v7x6wto", "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/2l/c2likoimxa3yi6j6dwntng3kzuvmiu64h4ync66w4mrx4v7x6wto.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": 1336753, "tid": 1381210,
    "ts": 1514656755454.603, "dur": 27.730,
    "args": {
      "External id": 2483471,"kernel_hash": "c2mqya6ilpprpgchmqenn6k5mpg3mienn3ill2mhoh2uq42xagjg", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2m/c2mqya6ilpprpgchmqenn6k5mpg3mienn3ill2mhoh2uq42xagjg.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": 1336753, "tid": 1381210,
    "ts": 1514656755499.193, "dur": 20.529,
    "args": {
      "External id": 2483472,"kernel_hash": "cstbyq2jrfkh2pvulirtdouqrjc47q3sh3z3auux63nbg6d6xj43", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/st/cstbyq2jrfkh2pvulirtdouqrjc47q3sh3z3auux63nbg6d6xj43.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": 1336753, "tid": 1381210,
    "ts": 1514656755579.531, "dur": 13.552,
    "args": {
      "External id": 2483473,"Record function id": 0, "Ev Idx": 272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656755581.800, "dur": 10.444,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755584.936, "dur": 6.575,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755588.499, "dur": 2.862,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755596.647, "dur": 33.755,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755597.743, "dur": 25.646,
    "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": 1336753, "tid": 1381210, "ts": 1514656755597.743,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 1336753, "tid": 1381210,
    "ts": 1514656755599.454, "dur": 7.967,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755604.073, "dur": 1.268,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755608.003, "dur": 6.758,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755613.493, "dur": 0.461,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755615.119, "dur": 2.395,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755616.327, "dur": 0.473,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755617.891, "dur": 4.851,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755621.515, "dur": 0.606,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755633.893, "dur": 5.244,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755634.697, "dur": 0.990,
    "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": 1336753, "tid": 1381210, "ts": 1514656755634.697,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SwiGLULinearFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514656755642.982, "dur": 554.171,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755644.008, "dur": 539.755,
    "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": 1336753, "tid": 1381210, "ts": 1514656755644.008,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656755681.868, "dur": 12.436,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755689.488, "dur": 4.481,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755698.944, "dur": 5.727,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755701.147, "dur": 2.584,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755702.571, "dur": 0.937,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755710.564, "dur": 122.212,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755711.456, "dur": 8.106,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755712.145, "dur": 6.237,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755714.930, "dur": 3.314,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755720.979, "dur": 111.116,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755722.881, "dur": 108.060,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755838.139, "dur": 6.832,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755842.151, "dur": 2.703,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755876.764, "dur": 4.311,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755882.027, "dur": 3.386,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755888.433, "dur": 1.369,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755925.945, "dur": 2.187,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755926.759, "dur": 1.247,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755959.264, "dur": 198.989,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755966.510, "dur": 9.421,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755972.372, "dur": 1.236,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656755978.117, "dur": 47.058,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756020.542, "dur": 3.308,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756027.197, "dur": 4.320,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756030.669, "dur": 0.474,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756032.160, "dur": 4.846,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756035.796, "dur": 0.665,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756041.311, "dur": 4.758,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756045.026, "dur": 0.777,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756049.206, "dur": 7.603,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756054.519, "dur": 2.093,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756057.427, "dur": 3.812,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756060.514, "dur": 0.414,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756061.916, "dur": 2.187,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756062.474, "dur": 1.552,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756066.473, "dur": 77.592,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756145.483, "dur": 3.111,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756151.526, "dur": 2.161,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756152.819, "dur": 0.411,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756156.050, "dur": 1.018,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756210.190, "dur": 9.816,
    "args": {
      "External id": 2483531,"Record function id": 0, "Ev Idx": 330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656756212.378, "dur": 6.912,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756214.895, "dur": 3.501,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756216.056, "dur": 2.202,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756224.109, "dur": 9.233,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756224.855, "dur": 6.000,
    "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": 1336753, "tid": 1381210, "ts": 1514656756224.855,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656756228.899, "dur": 1.758,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756229.652, "dur": 0.883,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756236.838, "dur": 170.687,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756237.652, "dur": 162.582,
    "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": 1336753, "tid": 1381210, "ts": 1514656756237.652,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656756243.795, "dur": 4.110,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756245.392, "dur": 1.948,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756246.598, "dur": 0.531,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756249.013, "dur": 56.932,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756307.181, "dur": 5.645,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756308.062, "dur": 4.079,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756311.058, "dur": 0.875,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756314.348, "dur": 5.173,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756315.266, "dur": 3.791,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756318.517, "dur": 0.394,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756320.188, "dur": 79.235,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756412.337, "dur": 6.731,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756413.692, "dur": 3.642,
    "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": 1336753, "tid": 1381210, "ts": 1514656756413.692,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656756415.010, "dur": 2.168,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756415.730, "dur": 1.337,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756422.597, "dur": 11.534,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756425.877, "dur": 5.892,
    "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": 1336753, "tid": 1381210, "ts": 1514656756425.877,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656756426.614, "dur": 4.951,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756427.374, "dur": 3.698,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756430.579, "dur": 0.389,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756438.195, "dur": 5.007,
    "args": {
      "External id": 2483561,"Record function id": 0, "Ev Idx": 360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656756439.512, "dur": 3.124,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756440.525, "dur": 1.827,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756441.181, "dur": 1.068,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756448.421, "dur": 7.193,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756449.345, "dur": 4.477,
    "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": 1336753, "tid": 1381210, "ts": 1514656756449.345,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656756450.136, "dur": 3.539,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756452.723, "dur": 0.842,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756458.829, "dur": 159.564,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756459.681, "dur": 149.929,
    "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": 1336753, "tid": 1381210, "ts": 1514656756459.681,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656756462.283, "dur": 2.623,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756463.196, "dur": 1.330,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756463.942, "dur": 0.479,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756467.926, "dur": 44.780,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756514.200, "dur": 6.011,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756514.965, "dur": 4.728,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756518.698, "dur": 0.876,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756521.271, "dur": 22.802,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756522.571, "dur": 20.628,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756525.971, "dur": 16.698,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756545.501, "dur": 63.233,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756624.415, "dur": 38.357,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756625.525, "dur": 5.897,
    "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": 1336753, "tid": 1381210, "ts": 1514656756625.525,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656756626.917, "dur": 4.352,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756629.866, "dur": 1.313,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756634.648, "dur": 25.355,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756666.620, "dur": 9.535,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756667.752, "dur": 6.316,
    "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": 1336753, "tid": 1381210, "ts": 1514656756667.752,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656756668.685, "dur": 5.169,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756671.483, "dur": 1.893,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756672.592, "dur": 0.658,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756680.432, "dur": 5.131,
    "args": {
      "External id": 2483592,"Record function id": 0, "Ev Idx": 391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656756681.657, "dur": 3.420,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756682.641, "dur": 1.922,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756683.239, "dur": 1.237,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756690.075, "dur": 468.972,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756691.479, "dur": 450.499,
    "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": 1336753, "tid": 1381210, "ts": 1514656756691.479,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 1336753, "tid": 1381210,
    "ts": 1514656756716.983, "dur": 32.229,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756718.453, "dur": 30.572,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756721.271, "dur": 5.181,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756723.483, "dur": 2.479,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756727.641, "dur": 20.929,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756760.112, "dur": 4.335,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756762.959, "dur": 1.334,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756768.215, "dur": 3.967,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756769.034, "dur": 3.057,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756786.035, "dur": 2.075,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656756800.544, "dur": 3.301,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757017.927, "dur": 4.508,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757028.264, "dur": 40.723,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757041.187, "dur": 1.216,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757075.191, "dur": 28.922,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757076.955, "dur": 26.954,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757081.119, "dur": 3.478,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757085.924, "dur": 17.489,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757110.922, "dur": 2.806,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757112.219, "dur": 1.379,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757121.686, "dur": 4.448,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757124.382, "dur": 1.641,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757128.088, "dur": 4.289,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757128.768, "dur": 3.512,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757172.905, "dur": 10.157,
    "args": {
      "External id": 2483622,"Record function id": 0, "Ev Idx": 421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656757175.365, "dur": 6.805,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757177.836, "dur": 3.443,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757179.219, "dur": 1.978,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757186.963, "dur": 9.806,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757187.916, "dur": 6.609,
    "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": 1336753, "tid": 1381210, "ts": 1514656757187.916,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656757191.958, "dur": 2.359,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757193.018, "dur": 1.168,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757202.451, "dur": 149.985,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757203.440, "dur": 144.224,
    "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": 1336753, "tid": 1381210, "ts": 1514656757203.440,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656757206.541, "dur": 4.539,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757207.998, "dur": 2.449,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757209.441, "dur": 0.807,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757212.077, "dur": 62.536,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757275.868, "dur": 7.321,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757278.533, "dur": 4.029,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757279.709, "dur": 2.706,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757284.931, "dur": 5.458,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757286.090, "dur": 3.854,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757289.305, "dur": 0.535,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757291.467, "dur": 55.366,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757357.050, "dur": 10.845,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757358.022, "dur": 8.011,
    "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": 1336753, "tid": 1381210, "ts": 1514656757358.022,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656757361.587, "dur": 4.288,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757362.629, "dur": 3.056,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757371.254, "dur": 9.495,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757372.110, "dur": 6.463,
    "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": 1336753, "tid": 1381210, "ts": 1514656757372.110,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656757372.949, "dur": 5.430,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757373.879, "dur": 4.005,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757377.209, "dur": 0.566,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757384.589, "dur": 8.157,
    "args": {
      "External id": 2483652,"Record function id": 0, "Ev Idx": 451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656757385.837, "dur": 6.361,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757386.950, "dur": 4.992,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757390.333, "dur": 1.476,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757395.866, "dur": 6.338,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757396.782, "dur": 3.661,
    "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": 1336753, "tid": 1381210, "ts": 1514656757396.782,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656757397.873, "dur": 2.428,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757398.880, "dur": 1.317,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757408.611, "dur": 342.522,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757409.713, "dur": 323.630,
    "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": 1336753, "tid": 1381210, "ts": 1514656757409.713,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336753, "tid": 1381210,
    "ts": 1514656757425.268, "dur": 9.787,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757431.639, "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": 1336753, "tid": 1381210,
    "ts": 1514656757436.977, "dur": 2.788,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757437.860, "dur": 1.673,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757440.961, "dur": 6.686,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757444.265, "dur": 3.205,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757487.359, "dur": 219.549,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757595.066, "dur": 4.923,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757601.934, "dur": 4.193,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757719.864, "dur": 4.000,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757726.812, "dur": 0.829,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757729.459, "dur": 1.042,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757760.490, "dur": 268.141,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757761.993, "dur": 256.177,
    "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": 1336753, "tid": 1381210, "ts": 1514656757761.993,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1336753, "tid": 1381210,
    "ts": 1514656757783.042, "dur": 52.825,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757789.878, "dur": 3.272,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757794.386, "dur": 40.791,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757845.318, "dur": 3.813,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656757846.896, "dur": 1.937,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758038.870, "dur": 168.626,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758040.823, "dur": 160.839,
    "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": 1336753, "tid": 1381210, "ts": 1514656758040.823,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1336753, "tid": 1381210,
    "ts": 1514656758054.923, "dur": 40.084,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758060.179, "dur": 3.356,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758064.567, "dur": 29.814,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758102.587, "dur": 4.271,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758104.254, "dur": 2.305,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758215.951, "dur": 14.795,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758217.962, "dur": 10.313,
    "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": 1336753, "tid": 1381210, "ts": 1514656758217.962,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656758220.154, "dur": 7.836,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758221.462, "dur": 6.311,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758234.245, "dur": 7.702,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758235.408, "dur": 4.681,
    "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": 1336753, "tid": 1381210, "ts": 1514656758235.408,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656758238.357, "dur": 1.597,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758238.920, "dur": 0.887,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758245.111, "dur": 7.226,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758246.089, "dur": 4.493,
    "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": 1336753, "tid": 1381210, "ts": 1514656758246.089,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656758248.846, "dur": 1.592,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758249.499, "dur": 0.857,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758258.001, "dur": 5.451,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758258.881, "dur": 2.680,
    "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": 1336753, "tid": 1381210, "ts": 1514656758258.881,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656758259.906, "dur": 1.497,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758260.437, "dur": 0.877,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758266.927, "dur": 234.767,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758267.632, "dur": 226.926,
    "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": 1336753, "tid": 1381210, "ts": 1514656758267.632,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656758271.300, "dur": 8.256,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758275.418, "dur": 3.544,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758277.138, "dur": 1.502,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758280.878, "dur": 68.392,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758350.881, "dur": 6.424,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758351.777, "dur": 4.913,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758355.124, "dur": 1.409,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758435.892, "dur": 5.280,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758439.224, "dur": 1.373,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758439.962, "dur": 0.486,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758441.969, "dur": 51.594,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758507.107, "dur": 5.780,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758507.962, "dur": 3.385,
    "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": 1336753, "tid": 1381210, "ts": 1514656758507.962,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656758509.233, "dur": 1.985,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758509.927, "dur": 1.147,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758516.280, "dur": 29.245,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758519.405, "dur": 6.139,
    "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": 1336753, "tid": 1381210, "ts": 1514656758519.405,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656758520.083, "dur": 5.241,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758520.674, "dur": 4.083,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758524.106, "dur": 0.544,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758554.102, "dur": 10.030,
    "args": {
      "External id": 2483726,"Record function id": 0, "Ev Idx": 525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656758556.002, "dur": 7.329,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758558.678, "dur": 4.185,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758560.001, "dur": 2.757,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758569.524, "dur": 9.347,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758570.446, "dur": 6.366,
    "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": 1336753, "tid": 1381210, "ts": 1514656758570.446,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656758571.631, "dur": 5.019,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758575.407, "dur": 1.156,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758582.213, "dur": 103.464,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758583.158, "dur": 95.715,
    "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": 1336753, "tid": 1381210, "ts": 1514656758583.158,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656758585.958, "dur": 4.555,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758586.520, "dur": 3.567,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758589.383, "dur": 0.595,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758594.003, "dur": 36.288,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758631.601, "dur": 2.398,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758632.113, "dur": 1.390,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758632.850, "dur": 0.547,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758635.265, "dur": 6.917,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758636.114, "dur": 5.628,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758639.414, "dur": 2.249,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758642.835, "dur": 35.227,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758689.869, "dur": 35.407,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758690.715, "dur": 5.250,
    "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": 1336753, "tid": 1381210, "ts": 1514656758690.715,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656758694.103, "dur": 1.717,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758694.594, "dur": 1.133,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758698.941, "dur": 23.826,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758729.265, "dur": 8.550,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758730.176, "dur": 5.737,
    "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": 1336753, "tid": 1381210, "ts": 1514656758730.176,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656758730.944, "dur": 4.768,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758731.575, "dur": 3.596,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758734.539, "dur": 0.517,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758741.756, "dur": 6.557,
    "args": {
      "External id": 2483757,"Record function id": 0, "Ev Idx": 556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656758745.144, "dur": 2.651,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758745.999, "dur": 1.507,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758746.466, "dur": 0.949,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758751.458, "dur": 5.230,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758752.442, "dur": 2.580,
    "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": 1336753, "tid": 1381210, "ts": 1514656758752.442,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656758753.228, "dur": 1.657,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758754.017, "dur": 0.782,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758759.895, "dur": 88.515,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758762.355, "dur": 79.283,
    "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": 1336753, "tid": 1381210, "ts": 1514656758762.355,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656758764.461, "dur": 1.943,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758764.878, "dur": 1.136,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758765.553, "dur": 0.353,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758766.948, "dur": 29.431,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758797.345, "dur": 7.349,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758797.861, "dur": 6.311,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758803.625, "dur": 0.440,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758805.866, "dur": 2.213,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758806.340, "dur": 1.374,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758807.116, "dur": 0.536,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758808.515, "dur": 32.385,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758852.657, "dur": 25.191,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758853.435, "dur": 5.062,
    "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": 1336753, "tid": 1381210, "ts": 1514656758853.435,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656758856.870, "dur": 1.488,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758857.230, "dur": 1.007,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758860.756, "dur": 15.202,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758881.727, "dur": 10.740,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758882.684, "dur": 7.760,
    "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": 1336753, "tid": 1381210, "ts": 1514656758882.684,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656758883.335, "dur": 6.863,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758886.283, "dur": 3.353,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758889.129, "dur": 0.380,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758896.239, "dur": 7.083,
    "args": {
      "External id": 2483788,"Record function id": 0, "Ev Idx": 587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656758897.368, "dur": 5.426,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758898.252, "dur": 4.262,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758900.838, "dur": 1.556,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758907.299, "dur": 402.439,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758908.300, "dur": 368.077,
    "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": 1336753, "tid": 1381210, "ts": 1514656758908.300,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656758938.976, "dur": 1.742,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758939.446, "dur": 1.099,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758956.385, "dur": 5.767,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656758971.236, "dur": 1.924,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759166.785, "dur": 2.863,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759174.188, "dur": 39.202,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759187.310, "dur": 0.871,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759219.120, "dur": 31.110,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759220.939, "dur": 29.082,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759224.732, "dur": 4.768,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759231.005, "dur": 18.534,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759254.756, "dur": 2.182,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759255.746, "dur": 1.095,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759265.938, "dur": 3.411,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759268.323, "dur": 0.920,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759286.777, "dur": 17.764,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759321.584, "dur": 8.373,
    "args": {
      "External id": 2483810,"Record function id": 0, "Ev Idx": 609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656759323.775, "dur": 5.358,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759325.440, "dur": 2.592,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759326.249, "dur": 1.669,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759336.026, "dur": 5.097,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759336.986, "dur": 1.324,
    "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": 1336753, "tid": 1381210, "ts": 1514656759336.986,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SwiGLULinearFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514656759344.866, "dur": 422.556,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759346.109, "dur": 410.355,
    "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": 1336753, "tid": 1381210, "ts": 1514656759346.109,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656759374.391, "dur": 7.949,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759379.324, "dur": 2.750,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759385.495, "dur": 5.522,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759386.415, "dur": 4.042,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759389.652, "dur": 0.620,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759396.126, "dur": 82.255,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759396.833, "dur": 4.307,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759397.494, "dur": 3.240,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759398.062, "dur": 2.568,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759402.251, "dur": 75.628,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759403.434, "dur": 73.530,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759481.768, "dur": 4.701,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759485.126, "dur": 1.191,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759516.708, "dur": 4.990,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759522.694, "dur": 1.830,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759525.135, "dur": 1.524,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759579.875, "dur": 2.507,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759580.501, "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": 1336753, "tid": 1381210,
    "ts": 1514656759608.521, "dur": 127.834,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759613.309, "dur": 4.336,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759616.004, "dur": 0.947,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759618.874, "dur": 10.368,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759626.065, "dur": 2.663,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759630.724, "dur": 3.461,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759633.266, "dur": 0.577,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759634.936, "dur": 4.262,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759638.106, "dur": 0.672,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759642.703, "dur": 1.831,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759643.800, "dur": 0.430,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759649.378, "dur": 5.006,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759652.269, "dur": 1.928,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759657.801, "dur": 1.616,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759658.553, "dur": 0.555,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759660.050, "dur": 5.179,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759662.756, "dur": 2.373,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759666.777, "dur": 55.254,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759723.761, "dur": 3.460,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759727.998, "dur": 4.130,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759731.293, "dur": 0.499,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759734.260, "dur": 0.954,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759778.204, "dur": 7.612,
    "args": {
      "External id": 2483858,"Record function id": 0, "Ev Idx": 657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656759780.037, "dur": 5.134,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759781.804, "dur": 2.576,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759782.639, "dur": 1.643,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759789.399, "dur": 5.758,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759790.131, "dur": 3.222,
    "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": 1336753, "tid": 1381210, "ts": 1514656759790.131,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656759791.539, "dur": 1.623,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759792.120, "dur": 0.898,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759798.570, "dur": 119.227,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759799.339, "dur": 112.251,
    "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": 1336753, "tid": 1381210, "ts": 1514656759799.339,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656759805.010, "dur": 5.600,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759806.216, "dur": 3.903,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759809.577, "dur": 0.388,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759811.633, "dur": 38.866,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759851.643, "dur": 5.498,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759852.439, "dur": 4.045,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759855.508, "dur": 0.797,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759858.420, "dur": 2.674,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759859.115, "dur": 1.541,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759859.965, "dur": 0.629,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759861.912, "dur": 48.671,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759922.524, "dur": 7.879,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759923.371, "dur": 5.620,
    "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": 1336753, "tid": 1381210, "ts": 1514656759923.371,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656759926.924, "dur": 1.914,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759927.729, "dur": 0.958,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759933.925, "dur": 9.582,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759937.199, "dur": 4.016,
    "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": 1336753, "tid": 1381210, "ts": 1514656759937.199,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656759938.405, "dur": 2.618,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759939.230, "dur": 1.326,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759939.994, "dur": 0.476,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759947.130, "dur": 4.620,
    "args": {
      "External id": 2483888,"Record function id": 0, "Ev Idx": 687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656759948.347, "dur": 2.897,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759949.296, "dur": 1.674,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759949.938, "dur": 0.940,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759954.838, "dur": 10.078,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759957.841, "dur": 5.022,
    "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": 1336753, "tid": 1381210, "ts": 1514656759957.841,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656759958.655, "dur": 4.060,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759961.847, "dur": 0.729,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759968.329, "dur": 185.391,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759969.000, "dur": 174.696,
    "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": 1336753, "tid": 1381210, "ts": 1514656759969.000,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656759970.622, "dur": 63.487,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656759971.223, "dur": 61.905,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760031.273, "dur": 1.143,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760037.536, "dur": 43.623,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760082.269, "dur": 5.002,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760083.020, "dur": 3.686,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760084.123, "dur": 2.414,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760088.370, "dur": 8.866,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760088.899, "dur": 7.772,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760094.129, "dur": 2.460,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760097.658, "dur": 45.122,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760160.303, "dur": 31.390,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760161.406, "dur": 4.002,
    "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": 1336753, "tid": 1381210, "ts": 1514656760161.406,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656760162.955, "dur": 2.306,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760163.713, "dur": 1.411,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760168.231, "dur": 20.749,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760195.355, "dur": 10.633,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760196.528, "dur": 7.437,
    "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": 1336753, "tid": 1381210, "ts": 1514656760196.528,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656760197.125, "dur": 6.636,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760198.134, "dur": 5.028,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760202.577, "dur": 0.484,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760210.138, "dur": 5.614,
    "args": {
      "External id": 2483919,"Record function id": 0, "Ev Idx": 718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656760211.348, "dur": 3.864,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760212.708, "dur": 2.183,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760213.494, "dur": 1.311,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760221.866, "dur": 435.096,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760223.119, "dur": 400.699,
    "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": 1336753, "tid": 1381210, "ts": 1514656760223.119,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 1336753, "tid": 1381210,
    "ts": 1514656760246.958, "dur": 35.246,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760248.439, "dur": 33.565,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760251.032, "dur": 7.621,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760255.329, "dur": 2.924,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760260.095, "dur": 21.489,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760292.285, "dur": 5.751,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760296.445, "dur": 1.433,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760301.612, "dur": 3.950,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760302.526, "dur": 2.928,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760317.149, "dur": 2.096,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760329.419, "dur": 1.745,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760490.736, "dur": 2.301,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760497.117, "dur": 29.722,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760504.799, "dur": 1.018,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760551.999, "dur": 34.220,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760553.748, "dur": 32.240,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760559.932, "dur": 4.813,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760566.059, "dur": 19.440,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760591.174, "dur": 5.433,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760594.901, "dur": 1.573,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760604.459, "dur": 1.940,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760605.098, "dur": 1.177,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760610.563, "dur": 3.974,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760611.330, "dur": 3.109,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760638.230, "dur": 17.137,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760669.393, "dur": 10.332,
    "args": {
      "External id": 2483950,"Record function id": 0, "Ev Idx": 749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656760671.486, "dur": 7.481,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760673.114, "dur": 4.942,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760676.261, "dur": 1.665,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760684.130, "dur": 6.632,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760685.052, "dur": 3.581,
    "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": 1336753, "tid": 1381210, "ts": 1514656760685.052,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656760686.602, "dur": 1.854,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760687.227, "dur": 1.117,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760694.147, "dur": 123.948,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760695.047, "dur": 116.940,
    "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": 1336753, "tid": 1381210, "ts": 1514656760695.047,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656760697.724, "dur": 7.809,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760700.760, "dur": 4.153,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760703.966, "dur": 0.767,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760706.666, "dur": 54.040,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760761.716, "dur": 5.254,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760762.422, "dur": 3.911,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760763.232, "dur": 2.943,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760769.985, "dur": 2.555,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760770.809, "dur": 1.280,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760771.603, "dur": 0.389,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760773.044, "dur": 38.117,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760822.753, "dur": 10.327,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760823.582, "dur": 7.126,
    "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": 1336753, "tid": 1381210, "ts": 1514656760823.582,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656760827.043, "dur": 3.510,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760827.748, "dur": 2.705,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760838.621, "dur": 8.411,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760841.519, "dur": 3.661,
    "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": 1336753, "tid": 1381210, "ts": 1514656760841.519,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656760842.631, "dur": 2.346,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760843.171, "dur": 1.279,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760843.982, "dur": 0.357,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760850.963, "dur": 4.564,
    "args": {
      "External id": 2483980,"Record function id": 0, "Ev Idx": 779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656760852.180, "dur": 2.852,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760852.971, "dur": 1.841,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760853.830, "dur": 0.861,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760858.798, "dur": 9.660,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760859.441, "dur": 7.151,
    "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": 1336753, "tid": 1381210, "ts": 1514656760859.441,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656760862.698, "dur": 3.750,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760865.212, "dur": 1.096,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760872.512, "dur": 314.146,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760873.644, "dur": 294.594,
    "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": 1336753, "tid": 1381210, "ts": 1514656760873.644,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336753, "tid": 1381210,
    "ts": 1514656760886.744, "dur": 5.493,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760888.885, "dur": 2.895,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760893.930, "dur": 5.012,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760897.075, "dur": 1.655,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760900.334, "dur": 4.710,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760901.643, "dur": 3.238,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656760929.065, "dur": 212.421,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761047.051, "dur": 4.072,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761053.155, "dur": 4.580,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761154.849, "dur": 3.855,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761161.880, "dur": 0.999,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761164.902, "dur": 0.571,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761195.722, "dur": 216.389,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761197.476, "dur": 208.202,
    "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": 1336753, "tid": 1381210, "ts": 1514656761197.476,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1336753, "tid": 1381210,
    "ts": 1514656761217.917, "dur": 43.383,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761220.563, "dur": 3.225,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761225.043, "dur": 35.611,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761270.174, "dur": 6.090,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761273.725, "dur": 2.265,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761419.234, "dur": 170.352,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761420.615, "dur": 161.800,
    "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": 1336753, "tid": 1381210, "ts": 1514656761420.615,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1336753, "tid": 1381210,
    "ts": 1514656761432.293, "dur": 30.467,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761434.015, "dur": 2.306,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761437.369, "dur": 24.733,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761469.062, "dur": 8.281,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761474.888, "dur": 2.156,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761598.108, "dur": 14.818,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761599.810, "dur": 10.351,
    "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": 1336753, "tid": 1381210, "ts": 1514656761599.810,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656761602.323, "dur": 7.573,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761603.815, "dur": 5.892,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761616.479, "dur": 10.624,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761620.154, "dur": 4.968,
    "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": 1336753, "tid": 1381210, "ts": 1514656761620.154,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656761620.919, "dur": 4.053,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761623.801, "dur": 1.037,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761632.303, "dur": 5.629,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761633.427, "dur": 2.554,
    "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": 1336753, "tid": 1381210, "ts": 1514656761633.427,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656761634.181, "dur": 1.651,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761634.842, "dur": 0.910,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761641.456, "dur": 11.314,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761642.314, "dur": 8.226,
    "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": 1336753, "tid": 1381210, "ts": 1514656761642.314,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656761643.096, "dur": 7.307,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761649.490, "dur": 0.817,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761656.137, "dur": 142.562,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761657.079, "dur": 134.198,
    "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": 1336753, "tid": 1381210, "ts": 1514656761657.079,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656761660.706, "dur": 6.014,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761662.579, "dur": 3.544,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761664.330, "dur": 1.513,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761668.019, "dur": 64.832,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761734.145, "dur": 6.672,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761737.762, "dur": 2.474,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761738.773, "dur": 1.302,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761742.118, "dur": 5.296,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761743.474, "dur": 3.476,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761746.461, "dur": 0.428,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761748.133, "dur": 42.309,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761803.222, "dur": 8.507,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761804.125, "dur": 6.124,
    "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": 1336753, "tid": 1381210, "ts": 1514656761804.125,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656761808.166, "dur": 1.939,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761808.882, "dur": 1.122,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761815.009, "dur": 6.022,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761815.718, "dur": 3.644,
    "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": 1336753, "tid": 1381210, "ts": 1514656761815.718,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656761816.359, "dur": 2.794,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761817.061, "dur": 1.639,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761818.058, "dur": 0.557,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761828.453, "dur": 11.912,
    "args": {
      "External id": 2484054,"Record function id": 0, "Ev Idx": 853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656761830.389, "dur": 9.333,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761832.651, "dur": 6.742,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761836.744, "dur": 2.500,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761843.404, "dur": 5.165,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761844.530, "dur": 2.356,
    "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": 1336753, "tid": 1381210, "ts": 1514656761844.530,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656761845.373, "dur": 1.382,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761845.984, "dur": 0.651,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761851.562, "dur": 89.308,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761852.535, "dur": 83.051,
    "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": 1336753, "tid": 1381210, "ts": 1514656761852.535,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656761854.436, "dur": 4.103,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761857.016, "dur": 1.148,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761857.657, "dur": 0.400,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761859.211, "dur": 27.007,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761887.288, "dur": 5.574,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761888.282, "dur": 4.095,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761891.377, "dur": 0.885,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761896.078, "dur": 4.896,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761896.959, "dur": 3.638,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761897.738, "dur": 2.771,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761901.771, "dur": 32.986,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761945.225, "dur": 35.320,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761946.045, "dur": 5.524,
    "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": 1336753, "tid": 1381210, "ts": 1514656761946.045,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656761947.191, "dur": 4.237,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761950.258, "dur": 1.080,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656761954.352, "dur": 22.341,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762021.160, "dur": 10.474,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762023.856, "dur": 5.710,
    "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": 1336753, "tid": 1381210, "ts": 1514656762023.856,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656762024.780, "dur": 4.574,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762026.242, "dur": 2.347,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762027.454, "dur": 0.847,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762037.954, "dur": 9.212,
    "args": {
      "External id": 2484085,"Record function id": 0, "Ev Idx": 884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656762039.321, "dur": 7.322,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762040.761, "dur": 5.529,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762044.325, "dur": 1.851,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762050.607, "dur": 7.232,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762051.743, "dur": 4.559,
    "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": 1336753, "tid": 1381210, "ts": 1514656762051.743,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656762052.780, "dur": 3.372,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762055.243, "dur": 0.819,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762063.251, "dur": 102.056,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762064.061, "dur": 93.315,
    "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": 1336753, "tid": 1381210, "ts": 1514656762064.061,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656762066.311, "dur": 4.118,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762066.813, "dur": 3.147,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762069.358, "dur": 0.489,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762071.526, "dur": 33.935,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762106.479, "dur": 3.645,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762107.033, "dur": 2.576,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762108.173, "dur": 1.318,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762111.036, "dur": 9.693,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762116.320, "dur": 3.743,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762119.497, "dur": 0.444,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762121.496, "dur": 35.160,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762169.489, "dur": 24.941,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762170.553, "dur": 3.402,
    "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": 1336753, "tid": 1381210, "ts": 1514656762170.553,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656762171.788, "dur": 2.014,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762172.588, "dur": 1.054,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762176.706, "dur": 15.687,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762198.139, "dur": 10.562,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762201.015, "dur": 5.667,
    "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": 1336753, "tid": 1381210, "ts": 1514656762201.015,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656762201.971, "dur": 4.475,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762202.799, "dur": 3.120,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762205.366, "dur": 0.418,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762212.539, "dur": 5.278,
    "args": {
      "External id": 2484116,"Record function id": 0, "Ev Idx": 915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656762213.921, "dur": 3.432,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762214.666, "dur": 2.379,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762215.562, "dur": 1.344,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762221.598, "dur": 385.513,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762222.897, "dur": 347.100,
    "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": 1336753, "tid": 1381210, "ts": 1514656762222.897,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656762264.922, "dur": 4.681,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762267.996, "dur": 1.445,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762283.205, "dur": 6.360,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762297.993, "dur": 1.959,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762443.642, "dur": 1.716,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762449.431, "dur": 33.495,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762458.442, "dur": 0.933,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762491.239, "dur": 34.604,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762493.002, "dur": 32.571,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762500.942, "dur": 5.442,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762507.904, "dur": 17.202,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762548.898, "dur": 3.512,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762550.153, "dur": 2.014,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762560.880, "dur": 1.964,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762561.700, "dur": 1.008,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762582.788, "dur": 20.386,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762618.825, "dur": 10.368,
    "args": {
      "External id": 2484138,"Record function id": 0, "Ev Idx": 937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656762622.801, "dur": 5.547,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762624.668, "dur": 2.665,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762625.490, "dur": 1.729,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762632.888, "dur": 5.539,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762634.006, "dur": 1.279,
    "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": 1336753, "tid": 1381210, "ts": 1514656762634.006,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SwiGLULinearFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514656762642.382, "dur": 482.767,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762643.744, "dur": 469.888,
    "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": 1336753, "tid": 1381210, "ts": 1514656762643.744,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656762673.567, "dur": 8.289,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762678.608, "dur": 2.997,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762685.018, "dur": 9.993,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762686.530, "dur": 7.854,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762690.561, "dur": 3.668,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762698.499, "dur": 97.512,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762699.328, "dur": 4.950,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762702.161, "dur": 1.650,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762703.093, "dur": 0.636,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762705.781, "dur": 89.755,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762707.316, "dur": 87.437,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762799.691, "dur": 2.467,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762800.788, "dur": 1.233,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762833.128, "dur": 5.376,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762841.683, "dur": 1.817,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762844.498, "dur": 1.815,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762876.995, "dur": 1.983,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762877.547, "dur": 1.272,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762903.686, "dur": 189.223,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762908.578, "dur": 8.540,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762913.295, "dur": 2.859,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762918.823, "dur": 7.281,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762924.859, "dur": 0.607,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762927.736, "dur": 1.513,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762928.510, "dur": 0.382,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762931.772, "dur": 1.984,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762932.813, "dur": 0.564,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762939.336, "dur": 2.044,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762940.410, "dur": 0.708,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762942.199, "dur": 6.530,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762947.000, "dur": 1.554,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762949.434, "dur": 4.390,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762953.078, "dur": 0.481,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762956.812, "dur": 3.464,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762957.310, "dur": 2.873,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656762961.519, "dur": 114.324,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763081.255, "dur": 1.674,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763083.718, "dur": 4.406,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763087.148, "dur": 0.474,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763090.768, "dur": 0.995,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763137.588, "dur": 9.265,
    "args": {
      "External id": 2484186,"Record function id": 0, "Ev Idx": 985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656763139.838, "dur": 6.359,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763141.826, "dur": 3.534,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763143.072, "dur": 2.201,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763150.740, "dur": 8.430,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763151.833, "dur": 5.133,
    "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": 1336753, "tid": 1381210, "ts": 1514656763151.833,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656763155.124, "dur": 1.642,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763155.766, "dur": 0.867,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763162.652, "dur": 181.236,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763163.605, "dur": 174.267,
    "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": 1336753, "tid": 1381210, "ts": 1514656763163.605,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656763167.200, "dur": 6.620,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763168.913, "dur": 4.350,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763172.318, "dur": 0.727,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763174.930, "dur": 62.687,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763238.844, "dur": 6.301,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763239.658, "dur": 4.686,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763243.071, "dur": 1.119,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763246.510, "dur": 7.378,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763247.727, "dur": 5.711,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763250.823, "dur": 2.505,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763254.751, "dur": 82.379,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763348.272, "dur": 8.393,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763349.483, "dur": 5.849,
    "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": 1336753, "tid": 1381210, "ts": 1514656763349.483,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656763351.024, "dur": 4.160,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763353.951, "dur": 1.105,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763359.916, "dur": 9.032,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763360.636, "dur": 6.372,
    "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": 1336753, "tid": 1381210, "ts": 1514656763360.636,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656763361.647, "dur": 5.160,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763362.256, "dur": 4.070,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763365.733, "dur": 0.429,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763372.942, "dur": 5.864,
    "args": {
      "External id": 2484216,"Record function id": 0, "Ev Idx": 1015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656763374.483, "dur": 3.792,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763375.822, "dur": 2.227,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763376.802, "dur": 1.160,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763383.448, "dur": 7.808,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763384.519, "dur": 5.023,
    "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": 1336753, "tid": 1381210, "ts": 1514656763384.519,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656763385.602, "dur": 3.785,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763388.531, "dur": 0.723,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763394.425, "dur": 209.763,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763466.780, "dur": 129.122,
    "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": 1336753, "tid": 1381210, "ts": 1514656763466.780,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656763471.131, "dur": 2.494,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763471.637, "dur": 1.472,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763472.347, "dur": 0.541,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763474.083, "dur": 37.659,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763512.987, "dur": 8.709,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763513.656, "dur": 7.515,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763516.775, "dur": 4.266,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763522.907, "dur": 3.985,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763523.701, "dur": 2.755,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763526.098, "dur": 0.269,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763541.951, "dur": 52.967,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763610.723, "dur": 36.765,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763611.726, "dur": 6.851,
    "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": 1336753, "tid": 1381210, "ts": 1514656763611.726,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656763615.266, "dur": 3.163,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763615.762, "dur": 2.523,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763621.473, "dur": 22.840,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763651.398, "dur": 8.828,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763652.239, "dur": 6.088,
    "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": 1336753, "tid": 1381210, "ts": 1514656763652.239,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656763653.127, "dur": 4.995,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763653.728, "dur": 3.875,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763656.828, "dur": 0.692,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763664.248, "dur": 7.737,
    "args": {
      "External id": 2484247,"Record function id": 0, "Ev Idx": 1046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656763665.591, "dur": 5.887,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763666.779, "dur": 4.360,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763669.670, "dur": 1.351,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763675.907, "dur": 455.071,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763677.197, "dur": 420.937,
    "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": 1336753, "tid": 1381210, "ts": 1514656763677.197,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 1336753, "tid": 1381210,
    "ts": 1514656763702.155, "dur": 32.421,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763703.379, "dur": 31.002,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763705.799, "dur": 5.562,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763708.065, "dur": 2.882,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763712.630, "dur": 21.144,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763747.542, "dur": 4.181,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763748.370, "dur": 3.215,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763755.597, "dur": 3.664,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763758.248, "dur": 0.922,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763771.314, "dur": 2.124,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763783.438, "dur": 1.957,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763942.798, "dur": 2.527,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763951.275, "dur": 30.649,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656763958.942, "dur": 0.915,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764022.650, "dur": 36.903,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764026.374, "dur": 32.922,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764032.375, "dur": 5.643,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764039.599, "dur": 19.209,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764064.900, "dur": 2.961,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764065.921, "dur": 1.795,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764075.787, "dur": 7.134,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764079.086, "dur": 3.728,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764084.907, "dur": 3.943,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764087.940, "dur": 0.818,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764112.231, "dur": 17.223,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764143.671, "dur": 8.405,
    "args": {
      "External id": 2484278,"Record function id": 0, "Ev Idx": 1077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656764145.686, "dur": 5.568,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764147.572, "dur": 2.911,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764148.498, "dur": 1.907,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764155.578, "dur": 8.792,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764156.377, "dur": 5.743,
    "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": 1336753, "tid": 1381210, "ts": 1514656764156.377,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656764160.011, "dur": 1.903,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764160.646, "dur": 1.144,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764167.814, "dur": 126.854,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764168.644, "dur": 120.465,
    "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": 1336753, "tid": 1381210, "ts": 1514656764168.644,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656764171.485, "dur": 6.924,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764172.948, "dur": 4.814,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764176.695, "dur": 0.874,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764179.621, "dur": 55.056,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764235.812, "dur": 7.398,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764236.379, "dur": 6.247,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764239.377, "dur": 3.106,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764244.635, "dur": 5.114,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764245.699, "dur": 3.630,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764248.612, "dur": 0.650,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764250.516, "dur": 37.708,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764299.273, "dur": 10.672,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764300.567, "dur": 7.249,
    "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": 1336753, "tid": 1381210, "ts": 1514656764300.567,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656764301.894, "dur": 5.744,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764304.510, "dur": 3.027,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764315.686, "dur": 9.184,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764316.729, "dur": 6.538,
    "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": 1336753, "tid": 1381210, "ts": 1514656764316.729,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656764317.571, "dur": 5.491,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764318.675, "dur": 3.895,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764321.885, "dur": 0.569,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764328.746, "dur": 4.690,
    "args": {
      "External id": 2484308,"Record function id": 0, "Ev Idx": 1107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656764329.957, "dur": 2.966,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764330.864, "dur": 1.785,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764331.536, "dur": 1.023,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764338.689, "dur": 5.560,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764339.333, "dur": 3.068,
    "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": 1336753, "tid": 1381210, "ts": 1514656764339.333,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656764340.457, "dur": 1.811,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764340.979, "dur": 1.154,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764349.952, "dur": 334.696,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764351.054, "dur": 314.164,
    "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": 1336753, "tid": 1381210, "ts": 1514656764351.054,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336753, "tid": 1381210,
    "ts": 1514656764364.787, "dur": 5.699,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764366.806, "dur": 3.262,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764372.629, "dur": 6.702,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764376.013, "dur": 3.106,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764411.870, "dur": 5.198,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764415.336, "dur": 1.541,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764444.875, "dur": 194.035,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764520.890, "dur": 4.965,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764546.208, "dur": 6.160,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764652.164, "dur": 4.068,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764659.046, "dur": 0.768,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764661.691, "dur": 0.760,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764693.850, "dur": 222.389,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764695.392, "dur": 214.466,
    "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": 1336753, "tid": 1381210, "ts": 1514656764695.392,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1336753, "tid": 1381210,
    "ts": 1514656764714.016, "dur": 43.469,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764716.651, "dur": 3.228,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764721.002, "dur": 35.894,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764766.431, "dur": 6.646,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764770.167, "dur": 2.558,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764923.174, "dur": 224.256,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764924.841, "dur": 215.465,
    "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": 1336753, "tid": 1381210, "ts": 1514656764924.841,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1336753, "tid": 1381210,
    "ts": 1514656764936.889, "dur": 91.008,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764938.865, "dur": 2.491,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656764955.109, "dur": 71.769,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765037.823, "dur": 7.590,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765042.004, "dur": 3.117,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765155.960, "dur": 19.720,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765157.381, "dur": 15.268,
    "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": 1336753, "tid": 1381210, "ts": 1514656765157.381,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656765159.811, "dur": 12.503,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765163.583, "dur": 8.570,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765179.352, "dur": 7.697,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765180.389, "dur": 5.002,
    "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": 1336753, "tid": 1381210, "ts": 1514656765180.389,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656765181.163, "dur": 4.087,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765184.125, "dur": 0.988,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765190.262, "dur": 5.318,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765191.070, "dur": 3.137,
    "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": 1336753, "tid": 1381210, "ts": 1514656765191.070,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656765192.046, "dur": 2.006,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765192.860, "dur": 1.075,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765198.778, "dur": 9.822,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765199.666, "dur": 7.254,
    "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": 1336753, "tid": 1381210, "ts": 1514656765199.666,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656765200.475, "dur": 6.297,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765205.619, "dur": 1.064,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765211.982, "dur": 145.969,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765212.938, "dur": 138.752,
    "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": 1336753, "tid": 1381210, "ts": 1514656765212.938,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656765216.692, "dur": 6.479,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765218.918, "dur": 3.643,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765220.835, "dur": 1.463,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765227.197, "dur": 66.151,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765294.534, "dur": 3.799,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765295.421, "dur": 2.218,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765296.419, "dur": 1.054,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765300.122, "dur": 5.468,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765301.357, "dur": 3.783,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765304.520, "dur": 0.508,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765308.207, "dur": 42.628,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765364.779, "dur": 6.317,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765365.920, "dur": 3.744,
    "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": 1336753, "tid": 1381210, "ts": 1514656765365.920,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656765367.626, "dur": 1.911,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765368.280, "dur": 1.121,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765374.628, "dur": 8.620,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765375.749, "dur": 5.648,
    "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": 1336753, "tid": 1381210, "ts": 1514656765375.749,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656765376.579, "dur": 4.632,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765377.304, "dur": 3.419,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765380.109, "dur": 0.532,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765391.002, "dur": 9.471,
    "args": {
      "External id": 2484382,"Record function id": 0, "Ev Idx": 1181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656765392.935, "dur": 6.747,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765395.329, "dur": 3.973,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765396.872, "dur": 2.343,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765403.530, "dur": 5.992,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765404.782, "dur": 3.158,
    "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": 1336753, "tid": 1381210, "ts": 1514656765404.782,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656765405.738, "dur": 2.058,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765406.703, "dur": 1.012,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765412.620, "dur": 93.323,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765413.332, "dur": 87.355,
    "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": 1336753, "tid": 1381210, "ts": 1514656765413.332,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656765417.973, "dur": 2.440,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765418.503, "dur": 1.496,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765419.330, "dur": 0.559,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765420.995, "dur": 27.505,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765449.503, "dur": 7.503,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765450.272, "dur": 6.168,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765455.569, "dur": 0.757,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765458.207, "dur": 4.512,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765458.961, "dur": 3.386,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765460.030, "dur": 2.240,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765463.381, "dur": 36.702,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765510.327, "dur": 55.185,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765511.256, "dur": 7.518,
    "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": 1336753, "tid": 1381210, "ts": 1514656765511.256,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656765514.681, "dur": 3.955,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765517.620, "dur": 0.908,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765521.434, "dur": 40.746,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765573.511, "dur": 8.916,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765574.825, "dur": 5.592,
    "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": 1336753, "tid": 1381210, "ts": 1514656765574.825,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656765575.975, "dur": 4.202,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765577.085, "dur": 2.470,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765578.827, "dur": 0.611,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765586.920, "dur": 9.159,
    "args": {
      "External id": 2484413,"Record function id": 0, "Ev Idx": 1212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656765588.209, "dur": 7.353,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765589.711, "dur": 5.568,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765593.501, "dur": 1.694,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765599.119, "dur": 11.590,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765600.139, "dur": 8.662,
    "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": 1336753, "tid": 1381210, "ts": 1514656765600.139,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656765601.222, "dur": 7.437,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765607.226, "dur": 1.346,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765613.758, "dur": 94.712,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765614.372, "dur": 87.266,
    "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": 1336753, "tid": 1381210, "ts": 1514656765614.372,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656765618.705, "dur": 2.431,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765619.187, "dur": 1.558,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765620.032, "dur": 0.583,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765621.665, "dur": 31.832,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765654.555, "dur": 6.691,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765655.282, "dur": 5.473,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765658.498, "dur": 2.130,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765662.250, "dur": 4.739,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765663.294, "dur": 3.250,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765665.993, "dur": 0.487,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765667.637, "dur": 33.363,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765712.807, "dur": 25.152,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765713.878, "dur": 3.222,
    "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": 1336753, "tid": 1381210, "ts": 1514656765713.878,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656765715.112, "dur": 1.842,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765715.957, "dur": 0.907,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765719.484, "dur": 16.715,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765743.826, "dur": 7.585,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765744.863, "dur": 4.501,
    "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": 1336753, "tid": 1381210, "ts": 1514656765744.863,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656765745.894, "dur": 3.263,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765746.710, "dur": 1.941,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765747.904, "dur": 0.643,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765757.832, "dur": 4.357,
    "args": {
      "External id": 2484444,"Record function id": 0, "Ev Idx": 1243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656765759.077, "dur": 2.667,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765759.861, "dur": 1.626,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765760.362, "dur": 1.063,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765768.009, "dur": 377.983,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765769.445, "dur": 345.696,
    "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": 1336753, "tid": 1381210, "ts": 1514656765769.445,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656765800.327, "dur": 1.962,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765800.989, "dur": 1.163,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765815.194, "dur": 5.214,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765828.630, "dur": 1.596,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765967.014, "dur": 1.739,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765972.573, "dur": 73.956,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656765981.295, "dur": 0.819,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766054.415, "dur": 36.532,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766058.744, "dur": 31.999,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766064.993, "dur": 6.043,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766072.452, "dur": 17.540,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766095.386, "dur": 2.553,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766096.298, "dur": 1.503,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766104.848, "dur": 3.668,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766107.480, "dur": 0.931,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766124.767, "dur": 17.325,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766157.653, "dur": 7.908,
    "args": {
      "External id": 2484466,"Record function id": 0, "Ev Idx": 1265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656766159.518, "dur": 5.376,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766161.390, "dur": 2.627,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766162.251, "dur": 1.655,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766169.189, "dur": 4.882,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766170.385, "dur": 1.211,
    "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": 1336753, "tid": 1381210, "ts": 1514656766170.385,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SwiGLULinearFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514656766177.545, "dur": 419.886,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766180.654, "dur": 406.190,
    "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": 1336753, "tid": 1381210, "ts": 1514656766180.654,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656766209.800, "dur": 7.670,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766214.400, "dur": 2.812,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766220.402, "dur": 6.406,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766221.851, "dur": 4.309,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766223.229, "dur": 2.792,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766229.902, "dur": 85.478,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766232.652, "dur": 4.637,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766233.554, "dur": 3.288,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766236.235, "dur": 0.530,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766238.509, "dur": 76.348,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766239.927, "dur": 74.031,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766318.544, "dur": 2.094,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766319.353, "dur": 1.139,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766351.102, "dur": 5.755,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766360.323, "dur": 1.954,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766363.108, "dur": 1.946,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766396.014, "dur": 2.280,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766397.119, "dur": 1.011,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766423.095, "dur": 142.435,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766427.868, "dur": 10.328,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766432.290, "dur": 5.071,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766439.346, "dur": 6.058,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766444.131, "dur": 0.458,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766446.713, "dur": 1.779,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766447.623, "dur": 0.518,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766450.895, "dur": 2.175,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766451.661, "dur": 0.759,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766456.305, "dur": 3.739,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766459.164, "dur": 0.610,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766460.763, "dur": 6.354,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766465.457, "dur": 1.491,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766469.792, "dur": 3.539,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766472.543, "dur": 0.520,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766473.795, "dur": 3.495,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766474.334, "dur": 2.858,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766478.423, "dur": 70.068,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766553.768, "dur": 1.671,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766556.324, "dur": 2.374,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766557.743, "dur": 0.571,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766563.332, "dur": 1.097,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766609.552, "dur": 9.039,
    "args": {
      "External id": 2484514,"Record function id": 0, "Ev Idx": 1313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656766611.648, "dur": 6.228,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766613.697, "dur": 3.328,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766614.752, "dur": 2.154,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766622.167, "dur": 5.698,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766623.031, "dur": 3.000,
    "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": 1336753, "tid": 1381210, "ts": 1514656766623.031,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656766624.368, "dur": 1.448,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766624.856, "dur": 0.873,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766631.524, "dur": 126.742,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766632.508, "dur": 119.495,
    "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": 1336753, "tid": 1381210, "ts": 1514656766632.508,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656766635.608, "dur": 5.803,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766636.807, "dur": 4.055,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766639.862, "dur": 0.840,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766642.547, "dur": 45.804,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766689.492, "dur": 5.325,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766690.320, "dur": 3.748,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766693.082, "dur": 0.844,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766696.333, "dur": 7.555,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766699.684, "dur": 3.727,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766700.813, "dur": 2.462,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766704.565, "dur": 46.617,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766762.722, "dur": 5.846,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766763.571, "dur": 3.908,
    "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": 1336753, "tid": 1381210, "ts": 1514656766763.571,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656766765.128, "dur": 2.216,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766766.022, "dur": 1.220,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766772.128, "dur": 11.173,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766772.980, "dur": 8.408,
    "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": 1336753, "tid": 1381210, "ts": 1514656766772.980,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656766776.177, "dur": 4.954,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766776.728, "dur": 3.902,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766780.045, "dur": 0.458,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766787.070, "dur": 5.063,
    "args": {
      "External id": 2484544,"Record function id": 0, "Ev Idx": 1343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656766788.497, "dur": 3.120,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766789.588, "dur": 1.743,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766790.072, "dur": 1.153,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766795.461, "dur": 7.697,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766796.363, "dur": 4.980,
    "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": 1336753, "tid": 1381210, "ts": 1514656766796.363,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656766797.331, "dur": 3.870,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766800.455, "dur": 0.618,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766806.238, "dur": 113.036,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766807.000, "dur": 105.697,
    "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": 1336753, "tid": 1381210, "ts": 1514656766807.000,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656766809.167, "dur": 4.427,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766809.724, "dur": 3.453,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766812.413, "dur": 0.660,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766814.163, "dur": 35.702,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766852.583, "dur": 5.442,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766853.351, "dur": 4.175,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766854.215, "dur": 3.166,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766858.922, "dur": 2.346,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766859.772, "dur": 1.085,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766860.407, "dur": 0.384,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766863.630, "dur": 48.188,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766925.554, "dur": 32.763,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766926.501, "dur": 5.672,
    "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": 1336753, "tid": 1381210, "ts": 1514656766926.501,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656766930.448, "dur": 1.599,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766930.925, "dur": 0.962,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766934.972, "dur": 20.559,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766962.238, "dur": 10.042,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766963.416, "dur": 6.499,
    "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": 1336753, "tid": 1381210, "ts": 1514656766963.416,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656766964.256, "dur": 5.456,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766965.276, "dur": 3.856,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766968.434, "dur": 0.580,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766976.080, "dur": 8.022,
    "args": {
      "External id": 2484575,"Record function id": 0, "Ev Idx": 1374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656766979.845, "dur": 3.693,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766980.846, "dur": 2.231,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656766981.444, "dur": 1.526,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767026.943, "dur": 437.373,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767043.091, "dur": 388.860,
    "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": 1336753, "tid": 1381210, "ts": 1514656767043.091,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 1336753, "tid": 1381210,
    "ts": 1514656767069.038, "dur": 38.962,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767070.350, "dur": 37.468,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767073.212, "dur": 7.229,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767075.814, "dur": 4.114,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767081.848, "dur": 25.271,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767120.820, "dur": 6.353,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767123.445, "dur": 3.628,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767130.735, "dur": 1.390,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767131.300, "dur": 0.745,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767145.758, "dur": 2.364,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767158.641, "dur": 3.090,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767320.299, "dur": 3.859,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767327.748, "dur": 32.496,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767336.960, "dur": 0.978,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767365.377, "dur": 27.621,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767367.211, "dur": 25.586,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767370.521, "dur": 3.805,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767377.176, "dur": 15.229,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767397.778, "dur": 4.912,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767401.506, "dur": 1.052,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767409.568, "dur": 4.357,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767410.414, "dur": 3.359,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767415.819, "dur": 4.184,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767418.997, "dur": 0.923,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767445.842, "dur": 16.858,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767474.391, "dur": 8.273,
    "args": {
      "External id": 2484606,"Record function id": 0, "Ev Idx": 1405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656767476.179, "dur": 5.756,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767478.097, "dur": 3.004,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767479.081, "dur": 1.896,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767486.253, "dur": 9.600,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767487.283, "dur": 6.375,
    "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": 1336753, "tid": 1381210, "ts": 1514656767487.283,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656767489.177, "dur": 4.300,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767492.219, "dur": 1.120,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767499.297, "dur": 154.653,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767500.013, "dur": 147.598,
    "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": 1336753, "tid": 1381210, "ts": 1514656767500.013,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656767503.009, "dur": 4.282,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767504.520, "dur": 2.252,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767505.685, "dur": 0.857,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767508.286, "dur": 77.036,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767589.708, "dur": 7.505,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767590.565, "dur": 5.773,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767593.674, "dur": 2.533,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767598.943, "dur": 2.579,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767599.974, "dur": 1.110,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767600.518, "dur": 0.464,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767604.758, "dur": 42.152,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767660.596, "dur": 11.600,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767661.925, "dur": 8.273,
    "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": 1336753, "tid": 1381210, "ts": 1514656767661.925,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656767663.358, "dur": 6.684,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767666.014, "dur": 3.924,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767675.446, "dur": 9.601,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767676.470, "dur": 6.781,
    "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": 1336753, "tid": 1381210, "ts": 1514656767676.470,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656767677.667, "dur": 5.373,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767678.644, "dur": 3.832,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767681.832, "dur": 0.532,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767689.005, "dur": 6.329,
    "args": {
      "External id": 2484636,"Record function id": 0, "Ev Idx": 1435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656767690.609, "dur": 4.227,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767692.005, "dur": 2.558,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767693.042, "dur": 1.410,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767698.676, "dur": 5.324,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767699.403, "dur": 2.987,
    "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": 1336753, "tid": 1381210, "ts": 1514656767699.403,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656767700.563, "dur": 1.683,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767701.015, "dur": 1.108,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767708.256, "dur": 318.991,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767711.155, "dur": 257.932,
    "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": 1336753, "tid": 1381210, "ts": 1514656767711.155,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336753, "tid": 1381210,
    "ts": 1514656767725.542, "dur": 8.833,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767730.180, "dur": 3.805,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767736.267, "dur": 4.490,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767738.731, "dur": 1.825,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767742.433, "dur": 2.855,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767743.248, "dur": 1.869,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767769.135, "dur": 175.172,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767850.412, "dur": 4.296,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767858.426, "dur": 3.924,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767957.127, "dur": 3.130,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767962.961, "dur": 0.972,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656767965.816, "dur": 0.688,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768037.904, "dur": 235.105,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768039.757, "dur": 226.742,
    "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": 1336753, "tid": 1381210, "ts": 1514656768039.757,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1336753, "tid": 1381210,
    "ts": 1514656768060.280, "dur": 57.324,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768065.402, "dur": 3.996,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768071.038, "dur": 46.116,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768127.022, "dur": 4.635,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768128.662, "dur": 2.681,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768279.756, "dur": 153.446,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768283.234, "dur": 143.939,
    "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": 1336753, "tid": 1381210, "ts": 1514656768283.234,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1336753, "tid": 1381210,
    "ts": 1514656768294.876, "dur": 32.480,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768296.769, "dur": 2.520,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768300.536, "dur": 26.329,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768334.400, "dur": 4.325,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768335.752, "dur": 2.620,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768439.262, "dur": 15.060,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768440.668, "dur": 11.476,
    "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": 1336753, "tid": 1381210, "ts": 1514656768440.668,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656768444.472, "dur": 7.450,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768445.717, "dur": 6.063,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768458.179, "dur": 5.943,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768459.250, "dur": 2.866,
    "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": 1336753, "tid": 1381210, "ts": 1514656768459.250,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656768460.348, "dur": 1.643,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768461.096, "dur": 0.781,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768467.106, "dur": 9.292,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768467.869, "dur": 7.308,
    "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": 1336753, "tid": 1381210, "ts": 1514656768467.869,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656768473.336, "dur": 1.706,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768474.071, "dur": 0.834,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768479.565, "dur": 5.199,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768480.576, "dur": 2.452,
    "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": 1336753, "tid": 1381210, "ts": 1514656768480.576,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656768481.296, "dur": 1.606,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768481.980, "dur": 0.826,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768488.173, "dur": 178.796,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768488.812, "dur": 171.041,
    "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": 1336753, "tid": 1381210, "ts": 1514656768488.812,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656768491.906, "dur": 14.316,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768496.270, "dur": 9.301,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768503.643, "dur": 1.653,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768507.665, "dur": 88.638,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768599.003, "dur": 5.306,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768600.167, "dur": 3.177,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768601.816, "dur": 1.378,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768606.017, "dur": 6.220,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768610.079, "dur": 1.732,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768611.030, "dur": 0.669,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768613.059, "dur": 45.904,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768673.676, "dur": 8.666,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768674.593, "dur": 6.228,
    "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": 1336753, "tid": 1381210, "ts": 1514656768674.593,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656768678.415, "dur": 2.267,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768679.014, "dur": 1.548,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768685.499, "dur": 8.874,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768686.252, "dur": 6.123,
    "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": 1336753, "tid": 1381210, "ts": 1514656768686.252,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656768689.194, "dur": 2.987,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768689.947, "dur": 1.748,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768691.075, "dur": 0.534,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768700.041, "dur": 9.146,
    "args": {
      "External id": 2484710,"Record function id": 0, "Ev Idx": 1509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656768701.523, "dur": 6.898,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768704.033, "dur": 4.029,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768705.699, "dur": 2.248,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768712.582, "dur": 9.735,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768715.532, "dur": 4.997,
    "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": 1336753, "tid": 1381210, "ts": 1514656768715.532,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656768716.863, "dur": 3.526,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768719.533, "dur": 0.741,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768725.424, "dur": 90.469,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768726.038, "dur": 84.574,
    "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": 1336753, "tid": 1381210, "ts": 1514656768726.038,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656768728.360, "dur": 2.864,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768729.239, "dur": 1.567,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768730.139, "dur": 0.552,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768732.073, "dur": 27.895,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768763.020, "dur": 3.575,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768763.645, "dur": 2.371,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768764.532, "dur": 1.365,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768767.716, "dur": 6.968,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768768.596, "dur": 5.420,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768771.264, "dur": 2.628,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768777.157, "dur": 32.665,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768822.299, "dur": 36.207,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768823.117, "dur": 7.154,
    "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": 1336753, "tid": 1381210, "ts": 1514656768823.117,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656768824.608, "dur": 5.502,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768828.725, "dur": 1.286,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768833.491, "dur": 22.880,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768862.106, "dur": 91.306,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768942.443, "dur": 8.564,
    "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": 1336753, "tid": 1381210, "ts": 1514656768942.443,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656768945.484, "dur": 5.310,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768946.034, "dur": 4.162,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768949.465, "dur": 0.608,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768957.462, "dur": 4.806,
    "args": {
      "External id": 2484741,"Record function id": 0, "Ev Idx": 1540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656768958.639, "dur": 3.090,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768959.653, "dur": 1.794,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768960.302, "dur": 0.999,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768965.225, "dur": 6.916,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768966.001, "dur": 4.756,
    "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": 1336753, "tid": 1381210, "ts": 1514656768966.001,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656768966.793, "dur": 3.816,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768969.530, "dur": 0.999,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768975.183, "dur": 138.791,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768975.795, "dur": 129.414,
    "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": 1336753, "tid": 1381210, "ts": 1514656768975.795,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656768977.762, "dur": 3.851,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768978.107, "dur": 3.062,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768980.445, "dur": 0.615,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656768982.292, "dur": 71.385,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769058.557, "dur": 3.407,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769059.334, "dur": 1.922,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769060.430, "dur": 0.720,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769063.245, "dur": 2.207,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769064.055, "dur": 1.000,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769064.627, "dur": 0.362,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769068.310, "dur": 36.169,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769120.140, "dur": 29.015,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769121.025, "dur": 5.910,
    "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": 1336753, "tid": 1381210, "ts": 1514656769121.025,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656769124.723, "dur": 2.054,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769125.395, "dur": 1.256,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769129.581, "dur": 17.648,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769152.609, "dur": 8.813,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769153.362, "dur": 5.855,
    "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": 1336753, "tid": 1381210, "ts": 1514656769153.362,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336753, "tid": 1381210,
    "ts": 1514656769154.092, "dur": 4.929,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769154.960, "dur": 3.545,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769157.789, "dur": 0.611,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769165.626, "dur": 5.302,
    "args": {
      "External id": 2484772,"Record function id": 0, "Ev Idx": 1571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656769166.807, "dur": 3.619,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769167.926, "dur": 2.045,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769168.398, "dur": 1.441,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769174.744, "dur": 376.961,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769177.791, "dur": 316.188,
    "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": 1336753, "tid": 1381210, "ts": 1514656769177.791,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1381210,
    "ts": 1514656769208.944, "dur": 1.731,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769209.448, "dur": 1.072,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769225.769, "dur": 5.916,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769240.927, "dur": 2.278,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769389.363, "dur": 2.072,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769395.329, "dur": 37.396,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769406.997, "dur": 0.956,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769438.282, "dur": 31.667,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769440.223, "dur": 29.532,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769444.089, "dur": 4.424,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769452.602, "dur": 16.600,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769474.206, "dur": 4.688,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769477.574, "dur": 1.197,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769485.570, "dur": 1.735,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769486.289, "dur": 0.913,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769506.746, "dur": 19.927,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769564.908, "dur": 12.253,
    "args": {
      "External id": 2484794,"Record function id": 0, "Ev Idx": 1593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656769567.210, "dur": 8.999,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769569.732, "dur": 5.409,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769572.864, "dur": 2.134,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769581.536, "dur": 2735.614,
    "args": {
      "External id": 2484798,"Record function id": 0, "Ev Idx": 1597
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.27)", "pid": 1336753, "tid": 1381210,
    "ts": 1514656769612.308, "dur": 963.242,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769637.732, "dur": 929.657,
    "args": {
      "External id": 2484800,"Record function id": 0, "Ev Idx": 1599
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.26)", "pid": 1336753, "tid": 1381210,
    "ts": 1514656769649.989, "dur": 900.569,
    "args": {
      "External id": 2484801,"Record function id": 0, "Ev Idx": 1600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514656769719.739, "dur": 7.607,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769742.823, "dur": 36.626,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769749.191, "dur": 3.947,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769754.402, "dur": 0.560,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769759.293, "dur": 0.460,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769760.429, "dur": 0.508,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769761.455, "dur": 0.568,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769766.111, "dur": 0.344,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769767.074, "dur": 0.324,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769768.007, "dur": 1.954,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769772.794, "dur": 2.205,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769805.830, "dur": 35.377,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769877.302, "dur": 146.539,
    "args": {
      "External id": 2484814,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514656769890.017, "dur": 6.506,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769901.972, "dur": 8.362,
    "args": {
      "External id": 2484816,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514656769905.612, "dur": 4.361,
    "args": {
      "External id": 2484817,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1381210,
    "ts": 1514656769908.275, "dur": 0.700,
    "args": {
      "External id": 2484818,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514656769917.246, "dur": 30.473,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656769920.710, "dur": 0.286,
    "args": {
      "External id": 2484820,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514656769921.941, "dur": 2.078,
    "args": {
      "External id": 2484821,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1381210,
    "ts": 1514656769924.745, "dur": 0.330,
    "args": {
      "External id": 2484822,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1381210,
    "ts": 1514656769927.570, "dur": 0.304,
    "args": {
      "External id": 2484823,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1381210,
    "ts": 1514656769930.315, "dur": 0.352,
    "args": {
      "External id": 2484824,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1381210,
    "ts": 1514656769931.334, "dur": 2.286,
    "args": {
      "External id": 2484825,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1381210,
    "ts": 1514656769935.973, "dur": 0.367,
    "args": {
      "External id": 2484826,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1381210,
    "ts": 1514656769939.226, "dur": 0.464,
    "args": {
      "External id": 2484827,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1381210,
    "ts": 1514656769942.852, "dur": 0.268,
    "args": {
      "External id": 2484828,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1381210,
    "ts": 1514656769957.359, "dur": 23.271,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656770082.725, "dur": 354.871,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656770117.319, "dur": 315.490,
    "args": {
      "External id": 2484831,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514656770127.085, "dur": 299.549,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656770461.782, "dur": 2.243,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514656770584.300, "dur": 1713.280,
    "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": 1336753, "tid": 1381210, "ts": 1514656770584.300,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1381210,
    "ts": 1514656770711.356, "dur": 110.815,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656770869.001, "dur": 39.946,
    "args": {
      "External id": 2484836,"kernel_hash": "c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/5t/c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr.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": 1336753, "tid": 1381210,
    "ts": 1514656770927.959, "dur": 47.797,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656771022.613, "dur": 39.506,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656771071.838, "dur": 45.433,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656771127.195, "dur": 27.502,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656771161.385, "dur": 42.486,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656771232.544, "dur": 24.992,
    "args": {
      "External id": 2484842,"kernel_hash": "cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/mm/cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k.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": 1336753, "tid": 1381210,
    "ts": 1514656771278.278, "dur": 30.941,
    "args": {
      "External id": 2484843,"kernel_hash": "c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2", "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/4b/c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2.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": 1336753, "tid": 1381210,
    "ts": 1514656771334.703, "dur": 21.839,
    "args": {
      "External id": 2484844,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514656771373.422, "dur": 20.080,
    "args": {
      "External id": 2484845,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514656771403.499, "dur": 32.866,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656771439.382, "dur": 31.816,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656771500.387, "dur": 189.406,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656771596.465, "dur": 8.241,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656771606.591, "dur": 2.429,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656771726.683, "dur": 27.922,
    "args": {
      "External id": 2484851,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514656771772.353, "dur": 15.688,
    "args": {
      "External id": 2484852,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514656771796.021, "dur": 39.397,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656771841.291, "dur": 34.482,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656771882.135, "dur": 22.100,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656771910.515, "dur": 28.527,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656771944.381, "dur": 21.405,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656771971.364, "dur": 68.293,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772068.255, "dur": 26.086,
    "args": {
      "External id": 2484859,"kernel_hash": "cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/as/cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf.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": 1336753, "tid": 1381210,
    "ts": 1514656772113.524, "dur": 33.487,
    "args": {
      "External id": 2484860,"kernel_hash": "cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od", "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/ny/cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od.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": 1336753, "tid": 1381210,
    "ts": 1514656772175.508, "dur": 23.981,
    "args": {
      "External id": 2484861,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514656772214.785, "dur": 15.293,
    "args": {
      "External id": 2484862,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514656772252.582, "dur": 18.357,
    "args": {
      "External id": 2484863,"kernel_hash": "c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/25/c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6.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": 1336753, "tid": 1381210,
    "ts": 1514656772339.043, "dur": 14.244,
    "args": {
      "External id": 2484864,"Record function id": 0, "Ev Idx": 1663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656772342.020, "dur": 10.511,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772346.214, "dur": 5.477,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772347.667, "dur": 3.935,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772357.419, "dur": 6.814,
    "args": {
      "External id": 2484868,"Record function id": 0, "Ev Idx": 1667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656772361.039, "dur": 2.722,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772361.774, "dur": 1.488,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772362.419, "dur": 0.768,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772367.666, "dur": 4.464,
    "args": {
      "External id": 2484872,"Record function id": 0, "Ev Idx": 1671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656772368.757, "dur": 2.921,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772369.522, "dur": 1.659,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772370.265, "dur": 0.778,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772375.374, "dur": 6.513,
    "args": {
      "External id": 2484876,"Record function id": 0, "Ev Idx": 1675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656772376.554, "dur": 4.930,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772377.235, "dur": 3.772,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772377.568, "dur": 3.305,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772388.492, "dur": 3.655,
    "args": {
      "External id": 2484880,"Record function id": 0, "Ev Idx": 1679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656772389.424, "dur": 2.286,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772389.907, "dur": 1.379,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772390.380, "dur": 0.806,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772395.283, "dur": 3.691,
    "args": {
      "External id": 2484884,"Record function id": 0, "Ev Idx": 1683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656772396.446, "dur": 2.113,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772396.917, "dur": 1.086,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772397.353, "dur": 0.586,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772402.232, "dur": 3.793,
    "args": {
      "External id": 2484888,"Record function id": 0, "Ev Idx": 1687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656772403.253, "dur": 2.297,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772403.999, "dur": 1.127,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772404.465, "dur": 0.596,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772409.152, "dur": 6.015,
    "args": {
      "External id": 2484892,"Record function id": 0, "Ev Idx": 1691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656772410.310, "dur": 4.419,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772410.772, "dur": 3.412,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772413.592, "dur": 0.525,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772419.548, "dur": 3.348,
    "args": {
      "External id": 2484896,"Record function id": 0, "Ev Idx": 1695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656772420.427, "dur": 2.051,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772420.892, "dur": 1.019,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772421.167, "dur": 0.682,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772426.578, "dur": 175663.130,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772428.098, "dur": 175650.024,
    "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": 1336753, "tid": 1381210, "ts": 1514656772428.098,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.27)", "pid": 1336753, "tid": 1381210,
    "ts": 1514656772459.059, "dur": 39.589,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772506.600, "dur": 86.074,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772601.330, "dur": 175468.017,
    "args": {
      "External id": 2484904,"Record function id": 0, "Ev Idx": 1703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514656772657.251, "dur": 7.598,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772676.710, "dur": 4.832,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772700.584, "dur": 174525.790,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772716.312, "dur": 174500.057,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772802.419, "dur": 5.476,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772820.464, "dur": 174350.516,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772822.692, "dur": 174347.352,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772825.988, "dur": 11.376,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656772839.017, "dur": 174326.625,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656947326.264, "dur": 9.844,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656947329.266, "dur": 6.522,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656947373.070, "dur": 352.176,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656947409.590, "dur": 310.637,
    "args": {
      "External id": 2484917,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514656947423.143, "dur": 290.957,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656947749.765, "dur": 2.150,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1381210,
    "ts": 1514656947813.575, "dur": 6.524,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656947873.108, "dur": 1.018,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656947890.154, "dur": 1.549,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656947907.695, "dur": 1.004,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656947920.223, "dur": 0.690,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656947929.836, "dur": 0.711,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656947942.121, "dur": 0.914,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656947956.675, "dur": 2.179,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656947970.140, "dur": 1.055,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948106.388, "dur": 2702.741,
    "args": {
      "External id": 2484929,"Record function id": 0, "Ev Idx": 1728
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.26)", "pid": 1336753, "tid": 1381210,
    "ts": 1514656948125.757, "dur": 1045.695,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948144.665, "dur": 310.238,
    "args": {
      "External id": 2484931,"Record function id": 0, "Ev Idx": 1730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1381210,
    "ts": 1514656948219.328, "dur": 4.313,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948226.605, "dur": 0.853,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948229.379, "dur": 1.138,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948232.248, "dur": 3.362,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948238.969, "dur": 1.198,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948241.536, "dur": 1.046,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948244.107, "dur": 2.264,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948247.875, "dur": 0.864,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948254.130, "dur": 0.950,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948256.386, "dur": 0.469,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948273.520, "dur": 153.516,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948291.338, "dur": 130.990,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948313.266, "dur": 12.670,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948329.009, "dur": 65.802,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948333.729, "dur": 60.694,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948336.799, "dur": 8.905,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948347.395, "dur": 46.340,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948546.512, "dur": 617.261,
    "args": {
      "External id": 2484949,"Record function id": 0, "Ev Idx": 1748
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.25)", "pid": 1336753, "tid": 1381210,
    "ts": 1514656948566.251, "dur": 584.733,
    "args": {
      "External id": 2484950,"Record function id": 0, "Ev Idx": 1749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514656948620.678, "dur": 5.809,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948642.403, "dur": 38.226,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948648.788, "dur": 3.770,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948653.761, "dur": 0.621,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948655.298, "dur": 0.784,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948660.327, "dur": 0.426,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948661.238, "dur": 0.397,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948662.262, "dur": 0.396,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948667.124, "dur": 2.462,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948670.273, "dur": 0.442,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948673.712, "dur": 1.812,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948702.399, "dur": 33.676,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948765.560, "dur": 102.896,
    "args": {
      "External id": 2484963,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514656948775.233, "dur": 4.813,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948784.875, "dur": 11.328,
    "args": {
      "External id": 2484965,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514656948788.699, "dur": 7.095,
    "args": {
      "External id": 2484966,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1381210,
    "ts": 1514656948794.062, "dur": 0.528,
    "args": {
      "External id": 2484967,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514656948802.673, "dur": 27.326,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948804.128, "dur": 0.562,
    "args": {
      "External id": 2484969,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514656948807.869, "dur": 2.200,
    "args": {
      "External id": 2484970,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1381210,
    "ts": 1514656948810.952, "dur": 0.312,
    "args": {
      "External id": 2484971,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1381210,
    "ts": 1514656948812.010, "dur": 2.407,
    "args": {
      "External id": 2484972,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1381210,
    "ts": 1514656948818.420, "dur": 0.291,
    "args": {
      "External id": 2484973,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1381210,
    "ts": 1514656948819.396, "dur": 0.593,
    "args": {
      "External id": 2484974,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1381210,
    "ts": 1514656948822.435, "dur": 0.337,
    "args": {
      "External id": 2484975,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1381210,
    "ts": 1514656948824.772, "dur": 0.282,
    "args": {
      "External id": 2484976,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1381210,
    "ts": 1514656948825.553, "dur": 0.310,
    "args": {
      "External id": 2484977,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1381210,
    "ts": 1514656948841.573, "dur": 19.267,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948910.288, "dur": 159.913,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656948940.029, "dur": 126.526,
    "args": {
      "External id": 2484980,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514656948950.773, "dur": 111.048,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656949093.526, "dur": 1.995,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514656949179.264, "dur": 1609.754,
    "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": 1336753, "tid": 1381210, "ts": 1514656949179.264,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1381210,
    "ts": 1514656949283.536, "dur": 98.787,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656949424.933, "dur": 39.194,
    "args": {
      "External id": 2484985,"kernel_hash": "c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/5t/c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr.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": 1336753, "tid": 1381210,
    "ts": 1514656949479.880, "dur": 46.181,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656949555.256, "dur": 36.656,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656949600.812, "dur": 44.199,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656949651.771, "dur": 26.560,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656949684.994, "dur": 41.565,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656949751.825, "dur": 24.556,
    "args": {
      "External id": 2484991,"kernel_hash": "cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/mm/cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k.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": 1336753, "tid": 1381210,
    "ts": 1514656949799.891, "dur": 29.075,
    "args": {
      "External id": 2484992,"kernel_hash": "c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2", "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/4b/c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2.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": 1336753, "tid": 1381210,
    "ts": 1514656949850.195, "dur": 21.122,
    "args": {
      "External id": 2484993,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514656949889.907, "dur": 16.118,
    "args": {
      "External id": 2484994,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514656949913.692, "dur": 29.499,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656949946.076, "dur": 33.125,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950045.649, "dur": 166.354,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950123.101, "dur": 5.426,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950130.267, "dur": 2.366,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950252.385, "dur": 25.995,
    "args": {
      "External id": 2485000,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514656950291.780, "dur": 15.651,
    "args": {
      "External id": 2485001,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514656950314.991, "dur": 39.852,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950360.510, "dur": 33.848,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950405.218, "dur": 21.914,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950430.833, "dur": 28.746,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950464.624, "dur": 21.633,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950491.419, "dur": 27.587,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950555.955, "dur": 38.046,
    "args": {
      "External id": 2485008,"kernel_hash": "cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/as/cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf.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": 1336753, "tid": 1381210,
    "ts": 1514656950621.201, "dur": 28.435,
    "args": {
      "External id": 2485009,"kernel_hash": "cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od", "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/ny/cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od.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": 1336753, "tid": 1381210,
    "ts": 1514656950666.141, "dur": 18.093,
    "args": {
      "External id": 2485010,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514656950708.806, "dur": 15.786,
    "args": {
      "External id": 2485011,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514656950740.273, "dur": 16.671,
    "args": {
      "External id": 2485012,"kernel_hash": "c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/25/c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6.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": 1336753, "tid": 1381210,
    "ts": 1514656950831.325, "dur": 14.848,
    "args": {
      "External id": 2485013,"Record function id": 0, "Ev Idx": 1812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656950834.508, "dur": 10.818,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950838.627, "dur": 5.661,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950839.960, "dur": 4.167,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950850.181, "dur": 4.383,
    "args": {
      "External id": 2485017,"Record function id": 0, "Ev Idx": 1816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656950851.499, "dur": 2.606,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950852.289, "dur": 1.300,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950852.752, "dur": 0.717,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950857.915, "dur": 5.852,
    "args": {
      "External id": 2485021,"Record function id": 0, "Ev Idx": 1820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656950858.950, "dur": 4.361,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950859.659, "dur": 3.126,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950860.165, "dur": 2.520,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950867.112, "dur": 4.199,
    "args": {
      "External id": 2485025,"Record function id": 0, "Ev Idx": 1824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656950868.344, "dur": 2.549,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950869.086, "dur": 1.382,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950869.572, "dur": 0.804,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950874.458, "dur": 3.673,
    "args": {
      "External id": 2485029,"Record function id": 0, "Ev Idx": 1828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656950875.635, "dur": 2.092,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950876.287, "dur": 0.962,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950876.573, "dur": 0.575,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950881.240, "dur": 5.344,
    "args": {
      "External id": 2485033,"Record function id": 0, "Ev Idx": 1832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656950883.921, "dur": 2.190,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950884.539, "dur": 1.150,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950884.977, "dur": 0.620,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950889.832, "dur": 3.432,
    "args": {
      "External id": 2485037,"Record function id": 0, "Ev Idx": 1836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656950890.865, "dur": 1.949,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950891.331, "dur": 1.070,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950891.617, "dur": 0.682,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950896.341, "dur": 5.804,
    "args": {
      "External id": 2485041,"Record function id": 0, "Ev Idx": 1840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656950897.372, "dur": 4.358,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950898.023, "dur": 3.231,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950900.292, "dur": 0.869,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950905.293, "dur": 3.617,
    "args": {
      "External id": 2485045,"Record function id": 0, "Ev Idx": 1844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656950906.249, "dur": 2.256,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950906.853, "dur": 1.163,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950907.162, "dur": 0.754,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950912.484, "dur": 40366.019,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656950913.796, "dur": 40356.312,
    "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": 1336753, "tid": 1381210, "ts": 1514656950913.796,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.26)", "pid": 1336753, "tid": 1381210,
    "ts": 1514656950940.013, "dur": 36.564,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656951020.384, "dur": 62.590,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656951092.053, "dur": 40170.320,
    "args": {
      "External id": 2485053,"Record function id": 0, "Ev Idx": 1852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514656951179.477, "dur": 6.720,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656951195.716, "dur": 7.014,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656951220.327, "dur": 39201.489,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656951233.354, "dur": 39179.688,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656951302.442, "dur": 13.368,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656951326.392, "dur": 39050.530,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656951329.034, "dur": 39047.177,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656951334.164, "dur": 5.018,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656951340.888, "dur": 39032.034,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656990506.017, "dur": 8.226,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656990508.756, "dur": 5.019,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656990553.902, "dur": 370.330,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656990587.784, "dur": 331.645,
    "args": {
      "External id": 2485066,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514656990601.567, "dur": 311.985,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656990945.319, "dur": 2.085,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1381210,
    "ts": 1514656991055.026, "dur": 7.302,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991106.071, "dur": 1.534,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991123.644, "dur": 3.112,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991140.470, "dur": 1.026,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991151.637, "dur": 0.853,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991164.226, "dur": 1.033,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991174.645, "dur": 3.098,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991187.786, "dur": 2.244,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991202.867, "dur": 1.138,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991293.884, "dur": 2787.578,
    "args": {
      "External id": 2485078,"Record function id": 0, "Ev Idx": 1877
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.25)", "pid": 1336753, "tid": 1381210,
    "ts": 1514656991311.832, "dur": 1065.125,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991326.616, "dur": 325.408,
    "args": {
      "External id": 2485080,"Record function id": 0, "Ev Idx": 1879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1381210,
    "ts": 1514656991397.798, "dur": 4.082,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991404.756, "dur": 0.845,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991407.169, "dur": 2.709,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991411.343, "dur": 0.736,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991415.277, "dur": 0.825,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991417.390, "dur": 0.926,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991421.417, "dur": 1.951,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991424.677, "dur": 0.812,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991428.759, "dur": 0.754,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991430.687, "dur": 0.704,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991447.553, "dur": 172.597,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991465.626, "dur": 149.901,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991483.568, "dur": 15.936,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991502.252, "dur": 81.038,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991507.583, "dur": 75.311,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991510.528, "dur": 4.863,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991517.181, "dur": 64.907,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991731.798, "dur": 636.990,
    "args": {
      "External id": 2485098,"Record function id": 0, "Ev Idx": 1897
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.24)", "pid": 1336753, "tid": 1381210,
    "ts": 1514656991747.357, "dur": 608.719,
    "args": {
      "External id": 2485099,"Record function id": 0, "Ev Idx": 1898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514656991801.334, "dur": 5.353,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991822.496, "dur": 35.915,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991828.579, "dur": 1.484,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991833.046, "dur": 0.367,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991834.173, "dur": 0.594,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991837.725, "dur": 2.275,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991840.705, "dur": 0.518,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991843.863, "dur": 2.692,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991849.284, "dur": 0.319,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991850.089, "dur": 0.512,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991852.752, "dur": 0.267,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991867.244, "dur": 31.495,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991927.835, "dur": 191.424,
    "args": {
      "External id": 2485112,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514656991937.044, "dur": 3.870,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991945.412, "dur": 13.655,
    "args": {
      "External id": 2485114,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514656991949.459, "dur": 9.196,
    "args": {
      "External id": 2485115,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1381210,
    "ts": 1514656991954.283, "dur": 3.254,
    "args": {
      "External id": 2485116,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514656991965.169, "dur": 105.300,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656991966.723, "dur": 0.529,
    "args": {
      "External id": 2485118,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514656991969.637, "dur": 0.525,
    "args": {
      "External id": 2485119,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1381210,
    "ts": 1514656992045.649, "dur": 2.930,
    "args": {
      "External id": 2485120,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1381210,
    "ts": 1514656992053.809, "dur": 0.424,
    "args": {
      "External id": 2485121,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1381210,
    "ts": 1514656992055.009, "dur": 0.354,
    "args": {
      "External id": 2485122,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1381210,
    "ts": 1514656992056.037, "dur": 1.827,
    "args": {
      "External id": 2485123,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1381210,
    "ts": 1514656992060.636, "dur": 0.518,
    "args": {
      "External id": 2485124,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1381210,
    "ts": 1514656992061.804, "dur": 0.618,
    "args": {
      "External id": 2485125,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1381210,
    "ts": 1514656992064.911, "dur": 0.385,
    "args": {
      "External id": 2485126,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1381210,
    "ts": 1514656992084.318, "dur": 26.539,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656992166.546, "dur": 120.541,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656992198.708, "dur": 85.128,
    "args": {
      "External id": 2485129,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514656992207.411, "dur": 72.261,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656992302.153, "dur": 1.779,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514656992387.122, "dur": 1673.351,
    "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": 1336753, "tid": 1381210, "ts": 1514656992387.122,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1381210,
    "ts": 1514656992493.369, "dur": 121.600,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656992656.874, "dur": 39.888,
    "args": {
      "External id": 2485134,"kernel_hash": "c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/5t/c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr.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": 1336753, "tid": 1381210,
    "ts": 1514656992712.936, "dur": 47.761,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656992772.618, "dur": 31.191,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656992810.038, "dur": 42.742,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656992859.044, "dur": 27.764,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656992900.652, "dur": 41.660,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656992969.854, "dur": 61.782,
    "args": {
      "External id": 2485140,"kernel_hash": "cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/mm/cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k.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": 1336753, "tid": 1381210,
    "ts": 1514656993058.571, "dur": 30.346,
    "args": {
      "External id": 2485141,"kernel_hash": "c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2", "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/4b/c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2.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": 1336753, "tid": 1381210,
    "ts": 1514656993113.091, "dur": 21.334,
    "args": {
      "External id": 2485142,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514656993151.857, "dur": 15.756,
    "args": {
      "External id": 2485143,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514656993176.539, "dur": 35.531,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656993215.619, "dur": 36.402,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656993279.192, "dur": 172.582,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656993360.993, "dur": 7.241,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656993369.878, "dur": 2.395,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656993482.626, "dur": 25.136,
    "args": {
      "External id": 2485149,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514656993521.155, "dur": 32.681,
    "args": {
      "External id": 2485150,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514656993564.609, "dur": 41.888,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656993612.527, "dur": 34.599,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656993655.586, "dur": 21.977,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656993681.995, "dur": 42.965,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656993736.555, "dur": 26.683,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656993770.456, "dur": 29.596,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656993826.304, "dur": 26.688,
    "args": {
      "External id": 2485157,"kernel_hash": "cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/as/cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf.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": 1336753, "tid": 1381210,
    "ts": 1514656993872.349, "dur": 27.116,
    "args": {
      "External id": 2485158,"kernel_hash": "cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od", "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/ny/cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od.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": 1336753, "tid": 1381210,
    "ts": 1514656993913.650, "dur": 20.249,
    "args": {
      "External id": 2485159,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514656993950.812, "dur": 15.049,
    "args": {
      "External id": 2485160,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514656993979.789, "dur": 51.595,
    "args": {
      "External id": 2485161,"kernel_hash": "c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/25/c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6.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": 1336753, "tid": 1381210,
    "ts": 1514656994103.418, "dur": 13.993,
    "args": {
      "External id": 2485162,"Record function id": 0, "Ev Idx": 1961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656994106.483, "dur": 10.066,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994110.614, "dur": 5.125,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994112.060, "dur": 3.597,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994121.337, "dur": 6.663,
    "args": {
      "External id": 2485166,"Record function id": 0, "Ev Idx": 1965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656994124.651, "dur": 2.875,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994125.766, "dur": 1.337,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994126.143, "dur": 0.886,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994131.301, "dur": 6.581,
    "args": {
      "External id": 2485170,"Record function id": 0, "Ev Idx": 1969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656994132.409, "dur": 5.071,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994132.861, "dur": 4.212,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994133.807, "dur": 3.193,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994141.176, "dur": 3.823,
    "args": {
      "External id": 2485174,"Record function id": 0, "Ev Idx": 1973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656994142.365, "dur": 2.235,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994142.865, "dur": 1.322,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994143.293, "dur": 0.830,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994148.382, "dur": 3.227,
    "args": {
      "External id": 2485178,"Record function id": 0, "Ev Idx": 1977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656994149.308, "dur": 1.879,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994149.798, "dur": 0.950,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994150.093, "dur": 0.594,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994154.742, "dur": 3.453,
    "args": {
      "External id": 2485182,"Record function id": 0, "Ev Idx": 1981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656994155.895, "dur": 1.908,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994156.355, "dur": 1.027,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994156.676, "dur": 0.628,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994161.647, "dur": 5.768,
    "args": {
      "External id": 2485186,"Record function id": 0, "Ev Idx": 1985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656994162.571, "dur": 4.445,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994163.067, "dur": 3.561,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994165.699, "dur": 0.853,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994170.682, "dur": 3.542,
    "args": {
      "External id": 2485190,"Record function id": 0, "Ev Idx": 1989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656994171.768, "dur": 2.043,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994172.371, "dur": 1.027,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994172.691, "dur": 0.629,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994177.357, "dur": 3.571,
    "args": {
      "External id": 2485194,"Record function id": 0, "Ev Idx": 1993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514656994178.594, "dur": 1.926,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994179.046, "dur": 1.074,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994179.305, "dur": 0.738,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994184.365, "dur": 37265.687,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994186.212, "dur": 37255.747,
    "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": 1336753, "tid": 1381210, "ts": 1514656994186.212,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.25)", "pid": 1336753, "tid": 1381210,
    "ts": 1514656994215.570, "dur": 36.369,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994259.270, "dur": 60.740,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994329.642, "dur": 37104.811,
    "args": {
      "External id": 2485202,"Record function id": 0, "Ev Idx": 2001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514656994415.196, "dur": 7.209,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994433.182, "dur": 6.460,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994453.076, "dur": 36177.719,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994465.889, "dur": 36156.164,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994520.224, "dur": 30.331,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994559.437, "dur": 36023.383,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994561.788, "dur": 36020.064,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994565.377, "dur": 7.051,
    "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": 1336753, "tid": 1381210,
    "ts": 1514656994574.126, "dur": 36004.475,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657030715.556, "dur": 7.684,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657030717.961, "dur": 4.967,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657030753.388, "dur": 397.170,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657030785.829, "dur": 359.638,
    "args": {
      "External id": 2485215,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657030797.080, "dur": 342.672,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657031175.947, "dur": 2.549,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1381210,
    "ts": 1514657031237.106, "dur": 6.125,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657031284.136, "dur": 1.335,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657031300.737, "dur": 3.814,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657031316.962, "dur": 0.971,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657031327.978, "dur": 0.951,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657031337.838, "dur": 0.710,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657031349.710, "dur": 2.799,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657031362.414, "dur": 1.950,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657031375.803, "dur": 0.767,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657031464.143, "dur": 2824.737,
    "args": {
      "External id": 2485227,"Record function id": 0, "Ev Idx": 2026
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.24)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657031481.831, "dur": 1145.182,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657031495.634, "dur": 330.933,
    "args": {
      "External id": 2485229,"Record function id": 0, "Ev Idx": 2028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1381210,
    "ts": 1514657031587.545, "dur": 4.674,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657031595.616, "dur": 0.869,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657031598.095, "dur": 2.952,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657031602.584, "dur": 0.869,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657031606.776, "dur": 0.942,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657031609.051, "dur": 0.973,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657031611.335, "dur": 1.969,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657031614.531, "dur": 0.737,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657031620.906, "dur": 0.672,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657031623.416, "dur": 0.614,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657031642.593, "dur": 153.819,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657031659.895, "dur": 132.407,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657031681.076, "dur": 15.252,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657031698.995, "dur": 65.278,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657031701.815, "dur": 62.138,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657031705.070, "dur": 6.533,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657031715.400, "dur": 47.706,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657031903.012, "dur": 715.598,
    "args": {
      "External id": 2485247,"Record function id": 0, "Ev Idx": 2046
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.23)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657031919.864, "dur": 685.646,
    "args": {
      "External id": 2485248,"Record function id": 0, "Ev Idx": 2047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657032084.499, "dur": 8.078,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657032115.184, "dur": 31.373,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657032120.523, "dur": 2.543,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657032124.304, "dur": 0.427,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657032125.246, "dur": 0.421,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657032128.125, "dur": 0.284,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657032130.383, "dur": 0.391,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657032131.990, "dur": 2.193,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657032136.624, "dur": 0.337,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657032138.245, "dur": 0.468,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657032139.965, "dur": 1.103,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657032156.565, "dur": 38.836,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657032233.634, "dur": 109.779,
    "args": {
      "External id": 2485261,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657032244.347, "dur": 3.802,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657032253.274, "dur": 9.741,
    "args": {
      "External id": 2485263,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657032257.468, "dur": 5.145,
    "args": {
      "External id": 2485264,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1381210,
    "ts": 1514657032260.910, "dur": 0.582,
    "args": {
      "External id": 2485265,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657032269.878, "dur": 31.000,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657032272.127, "dur": 0.356,
    "args": {
      "External id": 2485267,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657032274.144, "dur": 0.255,
    "args": {
      "External id": 2485268,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1381210,
    "ts": 1514657032275.704, "dur": 6.348,
    "args": {
      "External id": 2485269,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1381210,
    "ts": 1514657032283.484, "dur": 0.344,
    "args": {
      "External id": 2485270,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1381210,
    "ts": 1514657032285.318, "dur": 0.343,
    "args": {
      "External id": 2485271,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1381210,
    "ts": 1514657032288.362, "dur": 0.405,
    "args": {
      "External id": 2485272,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1381210,
    "ts": 1514657032290.315, "dur": 0.353,
    "args": {
      "External id": 2485273,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1381210,
    "ts": 1514657032292.038, "dur": 0.473,
    "args": {
      "External id": 2485274,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1381210,
    "ts": 1514657032294.905, "dur": 0.366,
    "args": {
      "External id": 2485275,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1381210,
    "ts": 1514657032312.400, "dur": 20.242,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657032391.134, "dur": 126.235,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657032426.595, "dur": 87.468,
    "args": {
      "External id": 2485278,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657032435.649, "dur": 73.700,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657032549.814, "dur": 2.711,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514657032634.742, "dur": 1633.744,
    "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": 1336753, "tid": 1381210, "ts": 1514657032634.742,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1381210,
    "ts": 1514657032737.591, "dur": 100.190,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657032878.452, "dur": 37.702,
    "args": {
      "External id": 2485283,"kernel_hash": "c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/5t/c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr.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": 1336753, "tid": 1381210,
    "ts": 1514657032931.530, "dur": 46.776,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657033030.238, "dur": 38.755,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657033075.377, "dur": 44.136,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657033125.858, "dur": 27.577,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657033160.104, "dur": 40.352,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657033228.745, "dur": 24.122,
    "args": {
      "External id": 2485289,"kernel_hash": "cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/mm/cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k.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": 1336753, "tid": 1381210,
    "ts": 1514657033274.092, "dur": 25.588,
    "args": {
      "External id": 2485290,"kernel_hash": "c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2", "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/4b/c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2.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": 1336753, "tid": 1381210,
    "ts": 1514657033318.309, "dur": 21.270,
    "args": {
      "External id": 2485291,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657033355.513, "dur": 16.022,
    "args": {
      "External id": 2485292,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657033380.505, "dur": 28.994,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657033412.564, "dur": 30.799,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657033476.521, "dur": 188.318,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657033572.501, "dur": 6.465,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657033581.103, "dur": 2.263,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657033698.307, "dur": 23.514,
    "args": {
      "External id": 2485298,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657033734.014, "dur": 14.819,
    "args": {
      "External id": 2485299,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657033756.667, "dur": 45.026,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657033807.457, "dur": 34.395,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657033854.985, "dur": 22.748,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657033881.993, "dur": 32.034,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657033919.185, "dur": 20.808,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657033946.894, "dur": 83.771,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034065.418, "dur": 30.203,
    "args": {
      "External id": 2485306,"kernel_hash": "cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/as/cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf.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": 1336753, "tid": 1381210,
    "ts": 1514657034116.961, "dur": 23.079,
    "args": {
      "External id": 2485307,"kernel_hash": "cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od", "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/ny/cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od.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": 1336753, "tid": 1381210,
    "ts": 1514657034154.153, "dur": 20.299,
    "args": {
      "External id": 2485308,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657034192.475, "dur": 16.242,
    "args": {
      "External id": 2485309,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657034221.083, "dur": 20.973,
    "args": {
      "External id": 2485310,"kernel_hash": "c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/25/c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6.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": 1336753, "tid": 1381210,
    "ts": 1514657034311.212, "dur": 14.918,
    "args": {
      "External id": 2485311,"Record function id": 0, "Ev Idx": 2110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657034314.940, "dur": 10.347,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034318.893, "dur": 5.459,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034320.272, "dur": 3.920,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034330.178, "dur": 6.465,
    "args": {
      "External id": 2485315,"Record function id": 0, "Ev Idx": 2114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657034331.506, "dur": 4.657,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034332.398, "dur": 3.187,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034332.884, "dur": 2.614,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034340.157, "dur": 4.979,
    "args": {
      "External id": 2485319,"Record function id": 0, "Ev Idx": 2118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657034341.804, "dur": 2.901,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034342.434, "dur": 1.810,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034343.232, "dur": 0.935,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034348.328, "dur": 4.205,
    "args": {
      "External id": 2485323,"Record function id": 0, "Ev Idx": 2122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657034349.698, "dur": 2.430,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034350.402, "dur": 1.301,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034351.090, "dur": 0.537,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034355.605, "dur": 4.628,
    "args": {
      "External id": 2485327,"Record function id": 0, "Ev Idx": 2126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657034357.416, "dur": 2.410,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034358.169, "dur": 1.179,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034358.647, "dur": 0.632,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034363.267, "dur": 3.582,
    "args": {
      "External id": 2485331,"Record function id": 0, "Ev Idx": 2130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657034364.520, "dur": 1.914,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034364.943, "dur": 1.033,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034365.263, "dur": 0.646,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034370.116, "dur": 5.407,
    "args": {
      "External id": 2485335,"Record function id": 0, "Ev Idx": 2134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657034371.303, "dur": 3.804,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034371.767, "dur": 2.928,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034373.941, "dur": 0.690,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034378.609, "dur": 3.678,
    "args": {
      "External id": 2485339,"Record function id": 0, "Ev Idx": 2138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657034379.861, "dur": 1.992,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034380.355, "dur": 1.092,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034380.631, "dur": 0.744,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034385.479, "dur": 6.206,
    "args": {
      "External id": 2485343,"Record function id": 0, "Ev Idx": 2142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657034386.889, "dur": 4.380,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034387.575, "dur": 3.271,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034388.122, "dur": 2.663,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034395.490, "dur": 37342.005,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034396.802, "dur": 37332.474,
    "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": 1336753, "tid": 1381210, "ts": 1514657034396.802,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.24)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657034422.967, "dur": 37.635,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034468.631, "dur": 73.710,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034553.884, "dur": 37167.862,
    "args": {
      "External id": 2485351,"Record function id": 0, "Ev Idx": 2150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657034641.905, "dur": 6.820,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034658.838, "dur": 4.754,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034677.849, "dur": 36152.952,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034691.710, "dur": 36131.184,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034747.756, "dur": 12.959,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034767.857, "dur": 36015.961,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034771.700, "dur": 36011.076,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034775.654, "dur": 4.746,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657034782.049, "dur": 35997.438,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657070921.195, "dur": 8.516,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657070924.396, "dur": 5.002,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657070958.628, "dur": 437.703,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657071024.344, "dur": 366.570,
    "args": {
      "External id": 2485364,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657071035.682, "dur": 348.867,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657071421.962, "dur": 2.635,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1381210,
    "ts": 1514657071485.022, "dur": 6.347,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657071550.845, "dur": 4.182,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657071573.784, "dur": 1.489,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657071588.377, "dur": 0.919,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657071602.156, "dur": 1.140,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657071614.728, "dur": 2.931,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657071628.794, "dur": 1.411,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657071642.395, "dur": 2.386,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657071658.343, "dur": 0.964,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657071752.641, "dur": 2669.672,
    "args": {
      "External id": 2485376,"Record function id": 0, "Ev Idx": 2175
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.23)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657071770.664, "dur": 1011.808,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657071784.354, "dur": 347.343,
    "args": {
      "External id": 2485378,"Record function id": 0, "Ev Idx": 2177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1381210,
    "ts": 1514657071859.118, "dur": 6.079,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657071868.463, "dur": 1.047,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657071871.294, "dur": 0.845,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657071874.161, "dur": 1.084,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657071878.901, "dur": 1.061,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657071881.416, "dur": 1.062,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657071884.128, "dur": 1.887,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657071887.397, "dur": 0.868,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657071891.676, "dur": 2.650,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657071895.981, "dur": 0.974,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657071913.979, "dur": 181.707,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657071930.354, "dur": 160.295,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657071948.766, "dur": 12.836,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657071964.631, "dur": 98.094,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657071967.213, "dur": 95.084,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657071970.696, "dur": 4.931,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657071979.238, "dur": 81.784,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657072213.387, "dur": 561.752,
    "args": {
      "External id": 2485396,"Record function id": 0, "Ev Idx": 2195
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.22)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657072228.214, "dur": 531.845,
    "args": {
      "External id": 2485397,"Record function id": 0, "Ev Idx": 2196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657072280.421, "dur": 5.415,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657072301.898, "dur": 32.654,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657072306.635, "dur": 2.292,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657072311.500, "dur": 0.668,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657072313.513, "dur": 0.788,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657072315.754, "dur": 2.918,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657072320.285, "dur": 0.409,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657072322.435, "dur": 0.246,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657072325.808, "dur": 0.280,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657072327.314, "dur": 0.304,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657072329.016, "dur": 0.356,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657072343.422, "dur": 28.526,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657072402.790, "dur": 103.890,
    "args": {
      "External id": 2485410,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657072412.676, "dur": 3.326,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657072421.103, "dur": 11.380,
    "args": {
      "External id": 2485412,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657072425.280, "dur": 6.779,
    "args": {
      "External id": 2485413,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1381210,
    "ts": 1514657072429.002, "dur": 1.672,
    "args": {
      "External id": 2485414,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657072439.298, "dur": 29.205,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657072441.359, "dur": 2.638,
    "args": {
      "External id": 2485416,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657072445.700, "dur": 0.539,
    "args": {
      "External id": 2485417,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1381210,
    "ts": 1514657072447.493, "dur": 0.559,
    "args": {
      "External id": 2485418,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1381210,
    "ts": 1514657072450.141, "dur": 0.433,
    "args": {
      "External id": 2485419,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1381210,
    "ts": 1514657072452.020, "dur": 0.557,
    "args": {
      "External id": 2485420,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1381210,
    "ts": 1514657072453.772, "dur": 1.918,
    "args": {
      "External id": 2485421,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1381210,
    "ts": 1514657072457.021, "dur": 0.488,
    "args": {
      "External id": 2485422,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1381210,
    "ts": 1514657072458.855, "dur": 0.481,
    "args": {
      "External id": 2485423,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1381210,
    "ts": 1514657072462.230, "dur": 2.131,
    "args": {
      "External id": 2485424,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1381210,
    "ts": 1514657072478.672, "dur": 20.873,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657072570.502, "dur": 118.722,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657072599.462, "dur": 86.204,
    "args": {
      "External id": 2485427,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657072608.863, "dur": 72.627,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657072710.532, "dur": 1.583,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514657072789.712, "dur": 1611.540,
    "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": 1336753, "tid": 1381210, "ts": 1514657072789.712,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1381210,
    "ts": 1514657072893.869, "dur": 135.742,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657073074.119, "dur": 38.258,
    "args": {
      "External id": 2485432,"kernel_hash": "c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/5t/c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr.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": 1336753, "tid": 1381210,
    "ts": 1514657073130.623, "dur": 52.418,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657073192.123, "dur": 31.745,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657073230.392, "dur": 43.390,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657073280.958, "dur": 27.324,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657073317.079, "dur": 41.159,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657073382.182, "dur": 21.289,
    "args": {
      "External id": 2485438,"kernel_hash": "cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/mm/cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k.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": 1336753, "tid": 1381210,
    "ts": 1514657073421.603, "dur": 25.360,
    "args": {
      "External id": 2485439,"kernel_hash": "c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2", "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/4b/c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2.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": 1336753, "tid": 1381210,
    "ts": 1514657073469.107, "dur": 19.800,
    "args": {
      "External id": 2485440,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657073502.829, "dur": 16.058,
    "args": {
      "External id": 2485441,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657073545.765, "dur": 34.601,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657073583.625, "dur": 32.098,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657073641.073, "dur": 168.894,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657073714.239, "dur": 6.191,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657073722.476, "dur": 2.086,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657073842.353, "dur": 27.489,
    "args": {
      "External id": 2485447,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657073883.651, "dur": 14.963,
    "args": {
      "External id": 2485448,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657073905.405, "dur": 37.176,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657073949.754, "dur": 67.875,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074037.808, "dur": 26.953,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074069.339, "dur": 30.118,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074105.615, "dur": 20.824,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074135.968, "dur": 40.085,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074204.565, "dur": 24.959,
    "args": {
      "External id": 2485455,"kernel_hash": "cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/as/cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf.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": 1336753, "tid": 1381210,
    "ts": 1514657074248.120, "dur": 27.503,
    "args": {
      "External id": 2485456,"kernel_hash": "cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od", "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/ny/cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od.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": 1336753, "tid": 1381210,
    "ts": 1514657074291.732, "dur": 17.026,
    "args": {
      "External id": 2485457,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657074328.782, "dur": 15.926,
    "args": {
      "External id": 2485458,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657074357.270, "dur": 15.477,
    "args": {
      "External id": 2485459,"kernel_hash": "c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/25/c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6.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": 1336753, "tid": 1381210,
    "ts": 1514657074447.105, "dur": 15.960,
    "args": {
      "External id": 2485460,"Record function id": 0, "Ev Idx": 2259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657074450.466, "dur": 11.697,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074454.040, "dur": 7.337,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074455.665, "dur": 5.600,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074467.182, "dur": 4.635,
    "args": {
      "External id": 2485464,"Record function id": 0, "Ev Idx": 2263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657074468.426, "dur": 2.967,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074469.337, "dur": 1.573,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074469.885, "dur": 0.907,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074475.268, "dur": 4.880,
    "args": {
      "External id": 2485468,"Record function id": 0, "Ev Idx": 2267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657074476.654, "dur": 3.090,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074477.620, "dur": 1.653,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074478.233, "dur": 0.870,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074483.342, "dur": 4.933,
    "args": {
      "External id": 2485472,"Record function id": 0, "Ev Idx": 2271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657074484.615, "dur": 3.264,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074485.768, "dur": 1.687,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074486.460, "dur": 0.876,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074491.612, "dur": 3.541,
    "args": {
      "External id": 2485476,"Record function id": 0, "Ev Idx": 2275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657074492.799, "dur": 1.954,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074493.424, "dur": 0.931,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074493.730, "dur": 0.527,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074498.285, "dur": 3.689,
    "args": {
      "External id": 2485480,"Record function id": 0, "Ev Idx": 2279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657074499.424, "dur": 2.142,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074500.028, "dur": 1.115,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074500.428, "dur": 0.613,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074505.225, "dur": 3.543,
    "args": {
      "External id": 2485484,"Record function id": 0, "Ev Idx": 2283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657074506.389, "dur": 1.951,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074506.814, "dur": 1.106,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074507.103, "dur": 0.715,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074511.854, "dur": 8.750,
    "args": {
      "External id": 2485488,"Record function id": 0, "Ev Idx": 2287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657074513.090, "dur": 7.109,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074513.801, "dur": 5.980,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074517.216, "dur": 2.505,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074523.719, "dur": 19.603,
    "args": {
      "External id": 2485492,"Record function id": 0, "Ev Idx": 2291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657074525.199, "dur": 16.662,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074525.723, "dur": 1.325,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074526.002, "dur": 0.945,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074549.268, "dur": 37501.637,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074551.149, "dur": 37490.304,
    "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": 1336753, "tid": 1381210, "ts": 1514657074551.149,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.23)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657074581.883, "dur": 35.000,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074624.648, "dur": 59.948,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074690.585, "dur": 37342.255,
    "args": {
      "External id": 2485500,"Record function id": 0, "Ev Idx": 2299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657074775.764, "dur": 6.401,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074791.197, "dur": 4.900,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074812.246, "dur": 36416.660,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074825.044, "dur": 36394.867,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074876.069, "dur": 12.981,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074896.086, "dur": 36283.012,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074899.712, "dur": 36278.697,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074903.773, "dur": 4.909,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657074910.384, "dur": 36264.473,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657111316.640, "dur": 8.071,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657111319.487, "dur": 4.832,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657111353.015, "dur": 344.583,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657111383.331, "dur": 309.404,
    "args": {
      "External id": 2485513,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657111394.401, "dur": 292.571,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657111721.527, "dur": 2.162,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1381210,
    "ts": 1514657111781.721, "dur": 6.639,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657111834.479, "dur": 1.161,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657111853.151, "dur": 1.303,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657111866.779, "dur": 0.926,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657111879.916, "dur": 0.858,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657111893.817, "dur": 0.667,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657111906.680, "dur": 0.953,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657111922.175, "dur": 1.734,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657111937.158, "dur": 0.572,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112067.285, "dur": 2673.462,
    "args": {
      "External id": 2485525,"Record function id": 0, "Ev Idx": 2324
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.22)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657112086.263, "dur": 1014.503,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112101.148, "dur": 317.646,
    "args": {
      "External id": 2485527,"Record function id": 0, "Ev Idx": 2326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1381210,
    "ts": 1514657112177.996, "dur": 4.649,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112185.540, "dur": 1.165,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112188.614, "dur": 1.068,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112191.486, "dur": 0.783,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112195.704, "dur": 0.925,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112198.358, "dur": 0.873,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112200.853, "dur": 1.616,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112203.818, "dur": 3.272,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112210.687, "dur": 1.042,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112213.137, "dur": 0.780,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112231.140, "dur": 153.062,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112246.392, "dur": 133.434,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112267.737, "dur": 11.652,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112282.720, "dur": 66.958,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112285.260, "dur": 64.040,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112289.438, "dur": 5.602,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112296.975, "dur": 51.651,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112496.040, "dur": 596.016,
    "args": {
      "External id": 2485545,"Record function id": 0, "Ev Idx": 2344
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.21)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657112511.883, "dur": 567.514,
    "args": {
      "External id": 2485546,"Record function id": 0, "Ev Idx": 2345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657112583.381, "dur": 6.117,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112605.125, "dur": 32.821,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112610.511, "dur": 1.893,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112614.791, "dur": 0.933,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112617.551, "dur": 2.893,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112622.195, "dur": 0.414,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112624.445, "dur": 0.368,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112626.524, "dur": 0.380,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112628.452, "dur": 0.430,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112630.804, "dur": 0.373,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112632.935, "dur": 0.462,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112652.129, "dur": 35.371,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112722.040, "dur": 101.819,
    "args": {
      "External id": 2485559,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657112732.181, "dur": 2.860,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112740.511, "dur": 11.899,
    "args": {
      "External id": 2485561,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657112744.558, "dur": 7.426,
    "args": {
      "External id": 2485562,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1381210,
    "ts": 1514657112748.134, "dur": 2.715,
    "args": {
      "External id": 2485563,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657112759.425, "dur": 25.248,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112761.861, "dur": 0.440,
    "args": {
      "External id": 2485565,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657112764.101, "dur": 0.605,
    "args": {
      "External id": 2485566,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1381210,
    "ts": 1514657112766.185, "dur": 0.381,
    "args": {
      "External id": 2485567,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1381210,
    "ts": 1514657112768.330, "dur": 0.378,
    "args": {
      "External id": 2485568,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1381210,
    "ts": 1514657112770.374, "dur": 0.433,
    "args": {
      "External id": 2485569,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1381210,
    "ts": 1514657112772.677, "dur": 0.328,
    "args": {
      "External id": 2485570,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1381210,
    "ts": 1514657112774.866, "dur": 0.290,
    "args": {
      "External id": 2485571,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1381210,
    "ts": 1514657112776.749, "dur": 2.419,
    "args": {
      "External id": 2485572,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1381210,
    "ts": 1514657112780.772, "dur": 0.411,
    "args": {
      "External id": 2485573,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1381210,
    "ts": 1514657112796.355, "dur": 19.938,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112868.013, "dur": 106.974,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657112889.581, "dur": 81.933,
    "args": {
      "External id": 2485576,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657112898.613, "dur": 68.542,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657113024.639, "dur": 2.911,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514657113107.642, "dur": 1614.959,
    "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": 1336753, "tid": 1381210, "ts": 1514657113107.642,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1381210,
    "ts": 1514657113210.102, "dur": 104.702,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657113351.936, "dur": 37.715,
    "args": {
      "External id": 2485581,"kernel_hash": "c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/5t/c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr.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": 1336753, "tid": 1381210,
    "ts": 1514657113407.332, "dur": 48.864,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657113464.739, "dur": 33.009,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657113504.781, "dur": 59.033,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657113574.368, "dur": 32.868,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657113616.082, "dur": 42.714,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657113683.085, "dur": 24.364,
    "args": {
      "External id": 2485587,"kernel_hash": "cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/mm/cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k.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": 1336753, "tid": 1381210,
    "ts": 1514657113733.106, "dur": 29.035,
    "args": {
      "External id": 2485588,"kernel_hash": "c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2", "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/4b/c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2.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": 1336753, "tid": 1381210,
    "ts": 1514657113783.272, "dur": 20.938,
    "args": {
      "External id": 2485589,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657113818.185, "dur": 16.452,
    "args": {
      "External id": 2485590,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657113847.163, "dur": 30.447,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657113880.880, "dur": 33.481,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657113940.343, "dur": 214.962,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114057.673, "dur": 6.349,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114066.526, "dur": 2.167,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114186.999, "dur": 24.778,
    "args": {
      "External id": 2485596,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657114223.148, "dur": 15.185,
    "args": {
      "External id": 2485597,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657114249.812, "dur": 40.695,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114297.366, "dur": 35.612,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114339.489, "dur": 22.845,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114368.082, "dur": 29.699,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114407.014, "dur": 22.568,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114436.953, "dur": 44.675,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114504.832, "dur": 36.850,
    "args": {
      "External id": 2485604,"kernel_hash": "cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/as/cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf.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": 1336753, "tid": 1381210,
    "ts": 1514657114561.489, "dur": 28.109,
    "args": {
      "External id": 2485605,"kernel_hash": "cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od", "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/ny/cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od.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": 1336753, "tid": 1381210,
    "ts": 1514657114604.527, "dur": 17.760,
    "args": {
      "External id": 2485606,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657114640.047, "dur": 16.099,
    "args": {
      "External id": 2485607,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657114675.874, "dur": 16.795,
    "args": {
      "External id": 2485608,"kernel_hash": "c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/25/c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6.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": 1336753, "tid": 1381210,
    "ts": 1514657114762.846, "dur": 15.366,
    "args": {
      "External id": 2485609,"Record function id": 0, "Ev Idx": 2408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657114766.045, "dur": 11.240,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114770.117, "dur": 6.371,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114771.960, "dur": 4.361,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114782.279, "dur": 5.896,
    "args": {
      "External id": 2485613,"Record function id": 0, "Ev Idx": 2412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657114783.723, "dur": 4.012,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114784.671, "dur": 2.584,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114786.115, "dur": 1.045,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114791.744, "dur": 4.781,
    "args": {
      "External id": 2485617,"Record function id": 0, "Ev Idx": 2416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657114792.969, "dur": 3.142,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114793.791, "dur": 1.881,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114794.586, "dur": 1.023,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114799.677, "dur": 3.807,
    "args": {
      "External id": 2485621,"Record function id": 0, "Ev Idx": 2420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657114800.972, "dur": 2.120,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114801.573, "dur": 1.126,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114801.912, "dur": 0.716,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114806.509, "dur": 3.569,
    "args": {
      "External id": 2485625,"Record function id": 0, "Ev Idx": 2424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657114807.781, "dur": 1.889,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114808.210, "dur": 1.024,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114808.530, "dur": 0.636,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114813.319, "dur": 3.871,
    "args": {
      "External id": 2485629,"Record function id": 0, "Ev Idx": 2428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657114814.556, "dur": 2.199,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114815.010, "dur": 1.289,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114815.466, "dur": 0.735,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114820.427, "dur": 6.074,
    "args": {
      "External id": 2485633,"Record function id": 0, "Ev Idx": 2432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657114821.806, "dur": 4.288,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114822.548, "dur": 3.079,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114822.859, "dur": 2.695,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114829.583, "dur": 4.410,
    "args": {
      "External id": 2485637,"Record function id": 0, "Ev Idx": 2436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657114831.003, "dur": 2.587,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114831.696, "dur": 1.452,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114831.991, "dur": 1.051,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114837.033, "dur": 4.477,
    "args": {
      "External id": 2485641,"Record function id": 0, "Ev Idx": 2440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657114838.556, "dur": 2.540,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114838.999, "dur": 1.657,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114840.022, "dur": 0.519,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114845.063, "dur": 36878.511,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114846.470, "dur": 36869.689,
    "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": 1336753, "tid": 1381210, "ts": 1514657114846.470,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.22)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657114876.487, "dur": 37.431,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657114921.452, "dur": 97.336,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657115028.182, "dur": 36680.408,
    "args": {
      "External id": 2485649,"Record function id": 0, "Ev Idx": 2448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657115124.438, "dur": 7.389,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657115141.668, "dur": 5.022,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657115161.921, "dur": 35705.288,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657115175.085, "dur": 35684.137,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657115235.169, "dur": 13.328,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657115255.912, "dur": 35560.750,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657115258.510, "dur": 35557.176,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657115263.023, "dur": 5.244,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657115270.098, "dur": 35542.324,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657150952.381, "dur": 9.315,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657150956.029, "dur": 5.341,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657151016.346, "dur": 358.751,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657151043.814, "dur": 326.444,
    "args": {
      "External id": 2485662,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657151054.968, "dur": 308.919,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657151395.474, "dur": 2.100,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1381210,
    "ts": 1514657151460.305, "dur": 8.784,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657151516.479, "dur": 1.307,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657151555.095, "dur": 1.662,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657151572.386, "dur": 0.924,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657151586.079, "dur": 2.688,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657151600.152, "dur": 1.169,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657151616.625, "dur": 0.837,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657151630.137, "dur": 1.814,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657151643.621, "dur": 2.719,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657151739.792, "dur": 2805.276,
    "args": {
      "External id": 2485674,"Record function id": 0, "Ev Idx": 2473
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.21)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657151758.364, "dur": 1062.845,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657151773.450, "dur": 394.881,
    "args": {
      "External id": 2485676,"Record function id": 0, "Ev Idx": 2475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1381210,
    "ts": 1514657151850.147, "dur": 4.011,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657151857.374, "dur": 0.965,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657151860.082, "dur": 1.128,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657151863.033, "dur": 0.949,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657151865.511, "dur": 1.121,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657151868.390, "dur": 1.085,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657151870.973, "dur": 3.709,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657151876.485, "dur": 0.964,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657151878.872, "dur": 1.207,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657151881.812, "dur": 0.767,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657151900.372, "dur": 228.729,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657151955.010, "dur": 168.156,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657151972.820, "dur": 47.578,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657152024.608, "dur": 67.937,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657152027.534, "dur": 64.607,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657152031.984, "dur": 6.787,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657152040.809, "dur": 50.787,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657152255.824, "dur": 556.508,
    "args": {
      "External id": 2485694,"Record function id": 0, "Ev Idx": 2493
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.20)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657152273.879, "dur": 525.082,
    "args": {
      "External id": 2485695,"Record function id": 0, "Ev Idx": 2494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657152332.240, "dur": 5.521,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657152353.548, "dur": 30.853,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657152358.555, "dur": 1.336,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657152362.468, "dur": 2.884,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657152366.949, "dur": 0.444,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657152369.231, "dur": 0.514,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657152371.642, "dur": 0.357,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657152373.611, "dur": 0.372,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657152375.627, "dur": 0.425,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657152377.813, "dur": 0.328,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657152379.918, "dur": 0.405,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657152394.422, "dur": 30.955,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657152456.309, "dur": 123.162,
    "args": {
      "External id": 2485708,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657152466.378, "dur": 5.309,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657152477.139, "dur": 10.596,
    "args": {
      "External id": 2485710,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657152481.459, "dur": 5.868,
    "args": {
      "External id": 2485711,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1381210,
    "ts": 1514657152485.571, "dur": 0.704,
    "args": {
      "External id": 2485712,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657152494.567, "dur": 26.193,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657152497.377, "dur": 0.683,
    "args": {
      "External id": 2485714,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657152499.865, "dur": 0.400,
    "args": {
      "External id": 2485715,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1381210,
    "ts": 1514657152502.069, "dur": 0.725,
    "args": {
      "External id": 2485716,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1381210,
    "ts": 1514657152504.431, "dur": 0.733,
    "args": {
      "External id": 2485717,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1381210,
    "ts": 1514657152506.952, "dur": 0.388,
    "args": {
      "External id": 2485718,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1381210,
    "ts": 1514657152508.823, "dur": 0.394,
    "args": {
      "External id": 2485719,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1381210,
    "ts": 1514657152510.916, "dur": 2.291,
    "args": {
      "External id": 2485720,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1381210,
    "ts": 1514657152514.920, "dur": 0.475,
    "args": {
      "External id": 2485721,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1381210,
    "ts": 1514657152517.009, "dur": 0.494,
    "args": {
      "External id": 2485722,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1381210,
    "ts": 1514657152546.878, "dur": 24.238,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657152624.790, "dur": 112.863,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657152648.130, "dur": 86.135,
    "args": {
      "External id": 2485725,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657152657.397, "dur": 72.423,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657152750.717, "dur": 1.875,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514657152828.133, "dur": 1681.619,
    "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": 1336753, "tid": 1381210, "ts": 1514657152828.133,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1381210,
    "ts": 1514657152932.643, "dur": 144.962,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657153123.014, "dur": 43.634,
    "args": {
      "External id": 2485730,"kernel_hash": "c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/5t/c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr.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": 1336753, "tid": 1381210,
    "ts": 1514657153186.511, "dur": 56.349,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657153252.737, "dur": 35.879,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657153295.590, "dur": 44.623,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657153350.054, "dur": 29.460,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657153386.478, "dur": 41.837,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657153450.193, "dur": 23.432,
    "args": {
      "External id": 2485736,"kernel_hash": "cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/mm/cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k.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": 1336753, "tid": 1381210,
    "ts": 1514657153490.851, "dur": 27.367,
    "args": {
      "External id": 2485737,"kernel_hash": "c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2", "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/4b/c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2.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": 1336753, "tid": 1381210,
    "ts": 1514657153553.741, "dur": 24.160,
    "args": {
      "External id": 2485738,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657153592.137, "dur": 17.396,
    "args": {
      "External id": 2485739,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657153622.304, "dur": 33.584,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657153659.580, "dur": 33.401,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657153719.258, "dur": 169.660,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657153792.798, "dur": 6.189,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657153801.064, "dur": 4.488,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657153941.717, "dur": 25.853,
    "args": {
      "External id": 2485745,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657153980.329, "dur": 51.546,
    "args": {
      "External id": 2485746,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657154044.757, "dur": 42.636,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154095.200, "dur": 36.543,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154139.700, "dur": 22.824,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154169.844, "dur": 45.503,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154227.800, "dur": 26.930,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154263.446, "dur": 32.165,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154314.799, "dur": 24.749,
    "args": {
      "External id": 2485753,"kernel_hash": "cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/as/cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf.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": 1336753, "tid": 1381210,
    "ts": 1514657154359.292, "dur": 25.077,
    "args": {
      "External id": 2485754,"kernel_hash": "cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od", "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/ny/cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od.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": 1336753, "tid": 1381210,
    "ts": 1514657154398.050, "dur": 21.288,
    "args": {
      "External id": 2485755,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657154434.110, "dur": 17.674,
    "args": {
      "External id": 2485756,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657154466.725, "dur": 15.987,
    "args": {
      "External id": 2485757,"kernel_hash": "c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/25/c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6.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": 1336753, "tid": 1381210,
    "ts": 1514657154567.638, "dur": 89.699,
    "args": {
      "External id": 2485758,"Record function id": 0, "Ev Idx": 2557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657154570.810, "dur": 85.307,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154574.991, "dur": 79.525,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154649.489, "dur": 4.841,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154662.128, "dur": 4.916,
    "args": {
      "External id": 2485762,"Record function id": 0, "Ev Idx": 2561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657154663.606, "dur": 2.973,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154664.342, "dur": 1.758,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154664.902, "dur": 1.112,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154670.344, "dur": 4.126,
    "args": {
      "External id": 2485766,"Record function id": 0, "Ev Idx": 2565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657154671.548, "dur": 2.518,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154672.089, "dur": 1.559,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154672.575, "dur": 0.990,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154677.688, "dur": 6.516,
    "args": {
      "External id": 2485770,"Record function id": 0, "Ev Idx": 2569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657154679.295, "dur": 4.456,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154679.953, "dur": 3.417,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154680.401, "dur": 2.888,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154687.300, "dur": 3.927,
    "args": {
      "External id": 2485774,"Record function id": 0, "Ev Idx": 2573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657154688.466, "dur": 2.327,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154688.899, "dur": 1.388,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154689.254, "dur": 0.966,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154694.463, "dur": 3.799,
    "args": {
      "External id": 2485778,"Record function id": 0, "Ev Idx": 2577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657154695.670, "dur": 2.185,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154696.345, "dur": 1.140,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154696.757, "dur": 0.652,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154701.821, "dur": 3.603,
    "args": {
      "External id": 2485782,"Record function id": 0, "Ev Idx": 2581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657154703.109, "dur": 1.876,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154703.608, "dur": 1.007,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154703.893, "dur": 0.631,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154708.724, "dur": 6.069,
    "args": {
      "External id": 2485786,"Record function id": 0, "Ev Idx": 2585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657154710.418, "dur": 3.966,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154710.853, "dur": 3.153,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154713.239, "dur": 0.656,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154718.323, "dur": 3.762,
    "args": {
      "External id": 2485790,"Record function id": 0, "Ev Idx": 2589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657154719.480, "dur": 2.172,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154720.046, "dur": 1.234,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154720.507, "dur": 0.693,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154726.152, "dur": 36983.955,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154727.492, "dur": 36974.862,
    "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": 1336753, "tid": 1381210, "ts": 1514657154727.492,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.21)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657154759.714, "dur": 36.101,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154803.524, "dur": 62.181,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154871.446, "dur": 36823.420,
    "args": {
      "External id": 2485798,"Record function id": 0, "Ev Idx": 2597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657154961.307, "dur": 7.144,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657154977.597, "dur": 42.052,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657155041.923, "dur": 35779.490,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657155055.756, "dur": 35757.185,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657155100.947, "dur": 15.626,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657155123.315, "dur": 35650.630,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657155126.030, "dur": 35647.108,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657155130.219, "dur": 5.907,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657155138.162, "dur": 35631.560,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657190906.401, "dur": 9.089,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657190909.924, "dur": 5.289,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657190945.272, "dur": 425.329,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657190971.234, "dur": 394.193,
    "args": {
      "External id": 2485811,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657191017.760, "dur": 341.230,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657191391.669, "dur": 2.187,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1381210,
    "ts": 1514657191453.677, "dur": 6.580,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657191504.347, "dur": 1.511,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657191522.920, "dur": 1.114,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657191557.019, "dur": 1.609,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657191573.275, "dur": 0.962,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657191587.081, "dur": 0.812,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657191601.523, "dur": 0.957,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657191617.711, "dur": 1.986,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657191632.388, "dur": 1.060,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657191726.017, "dur": 2727.533,
    "args": {
      "External id": 2485823,"Record function id": 0, "Ev Idx": 2622
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.20)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657191743.354, "dur": 1021.707,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657191757.988, "dur": 374.576,
    "args": {
      "External id": 2485825,"Record function id": 0, "Ev Idx": 2624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1381210,
    "ts": 1514657191832.654, "dur": 3.815,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657191839.607, "dur": 0.957,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657191842.439, "dur": 0.992,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657191845.381, "dur": 3.598,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657191850.499, "dur": 1.288,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657191853.571, "dur": 1.073,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657191856.539, "dur": 1.910,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657191860.125, "dur": 0.979,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657191862.704, "dur": 0.940,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657191865.168, "dur": 0.799,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657191886.262, "dur": 208.284,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657191901.813, "dur": 187.298,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657191926.307, "dur": 11.950,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657191941.540, "dur": 115.642,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657191944.247, "dur": 112.442,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657191948.434, "dur": 7.528,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657191957.997, "dur": 97.272,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657192210.620, "dur": 546.662,
    "args": {
      "External id": 2485843,"Record function id": 0, "Ev Idx": 2642
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.19)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657192228.366, "dur": 516.477,
    "args": {
      "External id": 2485844,"Record function id": 0, "Ev Idx": 2643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657192283.543, "dur": 5.370,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657192304.109, "dur": 33.511,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657192309.276, "dur": 1.679,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657192313.471, "dur": 0.610,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657192316.105, "dur": 0.477,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657192318.409, "dur": 0.417,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657192320.775, "dur": 0.502,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657192323.649, "dur": 0.449,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657192325.921, "dur": 2.493,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657192330.368, "dur": 0.484,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657192332.923, "dur": 0.522,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657192347.309, "dur": 30.970,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657192409.172, "dur": 100.397,
    "args": {
      "External id": 2485857,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657192418.725, "dur": 3.398,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657192427.476, "dur": 10.228,
    "args": {
      "External id": 2485859,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657192431.776, "dur": 5.537,
    "args": {
      "External id": 2485860,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1381210,
    "ts": 1514657192435.702, "dur": 0.526,
    "args": {
      "External id": 2485861,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657192444.605, "dur": 26.339,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657192446.995, "dur": 0.679,
    "args": {
      "External id": 2485863,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657192449.608, "dur": 0.609,
    "args": {
      "External id": 2485864,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1381210,
    "ts": 1514657192451.894, "dur": 0.550,
    "args": {
      "External id": 2485865,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1381210,
    "ts": 1514657192454.062, "dur": 2.403,
    "args": {
      "External id": 2485866,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1381210,
    "ts": 1514657192458.275, "dur": 0.392,
    "args": {
      "External id": 2485867,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1381210,
    "ts": 1514657192460.950, "dur": 0.358,
    "args": {
      "External id": 2485868,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1381210,
    "ts": 1514657192462.886, "dur": 0.317,
    "args": {
      "External id": 2485869,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1381210,
    "ts": 1514657192465.209, "dur": 0.405,
    "args": {
      "External id": 2485870,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1381210,
    "ts": 1514657192467.525, "dur": 0.386,
    "args": {
      "External id": 2485871,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1381210,
    "ts": 1514657192482.255, "dur": 20.546,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657192569.776, "dur": 112.527,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657192592.095, "dur": 86.785,
    "args": {
      "External id": 2485874,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657192601.826, "dur": 72.932,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657192695.481, "dur": 1.963,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514657192773.326, "dur": 1658.059,
    "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": 1336753, "tid": 1381210, "ts": 1514657192773.326,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1381210,
    "ts": 1514657192873.483, "dur": 97.646,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657193049.759, "dur": 40.526,
    "args": {
      "External id": 2485879,"kernel_hash": "c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/5t/c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr.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": 1336753, "tid": 1381210,
    "ts": 1514657193108.502, "dur": 55.477,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657193174.319, "dur": 34.167,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657193217.837, "dur": 45.689,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657193271.058, "dur": 29.899,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657193308.484, "dur": 40.830,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657193373.172, "dur": 23.151,
    "args": {
      "External id": 2485885,"kernel_hash": "cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/mm/cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k.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": 1336753, "tid": 1381210,
    "ts": 1514657193419.426, "dur": 27.935,
    "args": {
      "External id": 2485886,"kernel_hash": "c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2", "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/4b/c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2.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": 1336753, "tid": 1381210,
    "ts": 1514657193466.256, "dur": 20.240,
    "args": {
      "External id": 2485887,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657193500.323, "dur": 16.339,
    "args": {
      "External id": 2485888,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657193526.236, "dur": 50.484,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657193581.842, "dur": 35.668,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657193647.389, "dur": 180.472,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657193730.007, "dur": 5.683,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657193737.876, "dur": 5.965,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657193862.346, "dur": 25.683,
    "args": {
      "External id": 2485894,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657193899.267, "dur": 15.758,
    "args": {
      "External id": 2485895,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657193922.597, "dur": 34.431,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657193963.957, "dur": 74.111,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194051.654, "dur": 27.671,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194085.032, "dur": 31.806,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194123.933, "dur": 22.554,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194165.004, "dur": 45.033,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194235.008, "dur": 27.211,
    "args": {
      "External id": 2485902,"kernel_hash": "cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/as/cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf.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": 1336753, "tid": 1381210,
    "ts": 1514657194279.221, "dur": 27.214,
    "args": {
      "External id": 2485903,"kernel_hash": "cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od", "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/ny/cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od.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": 1336753, "tid": 1381210,
    "ts": 1514657194319.837, "dur": 18.363,
    "args": {
      "External id": 2485904,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657194357.279, "dur": 16.575,
    "args": {
      "External id": 2485905,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657194386.866, "dur": 16.818,
    "args": {
      "External id": 2485906,"kernel_hash": "c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/25/c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6.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": 1336753, "tid": 1381210,
    "ts": 1514657194475.442, "dur": 15.060,
    "args": {
      "External id": 2485907,"Record function id": 0, "Ev Idx": 2706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657194478.902, "dur": 10.656,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194482.921, "dur": 5.796,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194484.572, "dur": 4.037,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194494.344, "dur": 5.410,
    "args": {
      "External id": 2485911,"Record function id": 0, "Ev Idx": 2710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657194496.081, "dur": 3.238,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194496.995, "dur": 1.876,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194497.769, "dur": 0.967,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194503.164, "dur": 6.763,
    "args": {
      "External id": 2485915,"Record function id": 0, "Ev Idx": 2714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657194504.647, "dur": 4.824,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194505.216, "dur": 3.791,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194505.769, "dur": 3.176,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194513.300, "dur": 4.017,
    "args": {
      "External id": 2485919,"Record function id": 0, "Ev Idx": 2718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657194514.587, "dur": 2.322,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194515.196, "dur": 1.291,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194515.684, "dur": 0.731,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194520.633, "dur": 3.918,
    "args": {
      "External id": 2485923,"Record function id": 0, "Ev Idx": 2722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657194521.866, "dur": 2.294,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194522.696, "dur": 1.087,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194523.051, "dur": 0.667,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194543.154, "dur": 7.367,
    "args": {
      "External id": 2485927,"Record function id": 0, "Ev Idx": 2726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657194545.446, "dur": 4.372,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194546.451, "dur": 2.349,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194547.276, "dur": 1.275,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194555.072, "dur": 4.504,
    "args": {
      "External id": 2485931,"Record function id": 0, "Ev Idx": 2730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657194556.500, "dur": 2.669,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194557.478, "dur": 1.284,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194558.021, "dur": 0.638,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194562.921, "dur": 4.044,
    "args": {
      "External id": 2485935,"Record function id": 0, "Ev Idx": 2734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657194564.391, "dur": 2.147,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194564.916, "dur": 1.215,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194565.371, "dur": 0.694,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194570.633, "dur": 4.658,
    "args": {
      "External id": 2485939,"Record function id": 0, "Ev Idx": 2738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657194572.523, "dur": 2.328,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194573.112, "dur": 1.362,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194573.598, "dur": 0.787,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194579.164, "dur": 36888.853,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194580.760, "dur": 36879.718,
    "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": 1336753, "tid": 1381210, "ts": 1514657194580.760,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.20)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657194609.449, "dur": 37.525,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194654.781, "dur": 60.424,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194721.052, "dur": 36731.823,
    "args": {
      "External id": 2485947,"Record function id": 0, "Ev Idx": 2746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657194806.509, "dur": 6.498,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194822.057, "dur": 6.688,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194844.358, "dur": 35814.372,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194857.954, "dur": 35792.148,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194927.553, "dur": 13.378,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194947.226, "dur": 35662.972,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194949.732, "dur": 35659.681,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194954.110, "dur": 5.216,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657194961.155, "dur": 35644.659,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657230747.404, "dur": 8.845,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657230750.661, "dur": 5.251,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657230790.886, "dur": 358.988,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657230820.488, "dur": 324.230,
    "args": {
      "External id": 2485960,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657230830.910, "dur": 307.567,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657231171.071, "dur": 2.335,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1381210,
    "ts": 1514657231233.517, "dur": 6.370,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657231284.764, "dur": 1.347,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657231305.669, "dur": 2.685,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657231322.115, "dur": 0.900,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657231334.762, "dur": 1.000,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657231348.096, "dur": 1.037,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657231362.905, "dur": 3.062,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657231378.910, "dur": 2.123,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657231392.278, "dur": 0.962,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657231482.827, "dur": 2744.669,
    "args": {
      "External id": 2485972,"Record function id": 0, "Ev Idx": 2771
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.19)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657231501.405, "dur": 1007.536,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657231515.885, "dur": 338.319,
    "args": {
      "External id": 2485974,"Record function id": 0, "Ev Idx": 2773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1381210,
    "ts": 1514657231614.463, "dur": 4.604,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657231622.296, "dur": 1.277,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657231625.893, "dur": 3.070,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657231630.572, "dur": 0.840,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657231633.065, "dur": 0.877,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657231642.326, "dur": 1.000,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657231645.047, "dur": 2.068,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657231648.464, "dur": 1.152,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657231651.360, "dur": 0.881,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657231653.685, "dur": 0.934,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657231672.649, "dur": 149.038,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657231688.856, "dur": 128.286,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657231706.340, "dur": 14.169,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657231723.642, "dur": 64.233,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657231726.338, "dur": 61.207,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657231730.550, "dur": 5.587,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657231737.901, "dur": 49.005,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657231932.257, "dur": 568.666,
    "args": {
      "External id": 2485992,"Record function id": 0, "Ev Idx": 2791
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.18)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657231948.711, "dur": 539.427,
    "args": {
      "External id": 2485993,"Record function id": 0, "Ev Idx": 2792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657232042.773, "dur": 5.994,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657232064.449, "dur": 33.500,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657232070.039, "dur": 1.632,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657232074.091, "dur": 0.654,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657232076.558, "dur": 0.512,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657232079.138, "dur": 0.630,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657232081.788, "dur": 0.536,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657232084.312, "dur": 2.664,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657232088.819, "dur": 0.354,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657232090.900, "dur": 0.453,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657232093.142, "dur": 0.378,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657232107.900, "dur": 33.828,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657232173.016, "dur": 99.416,
    "args": {
      "External id": 2486006,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657232182.856, "dur": 3.268,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657232191.366, "dur": 10.308,
    "args": {
      "External id": 2486008,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657232195.732, "dur": 5.546,
    "args": {
      "External id": 2486009,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1381210,
    "ts": 1514657232199.439, "dur": 0.573,
    "args": {
      "External id": 2486010,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657232208.336, "dur": 26.699,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657232210.721, "dur": 0.388,
    "args": {
      "External id": 2486012,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657232213.122, "dur": 0.419,
    "args": {
      "External id": 2486013,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1381210,
    "ts": 1514657232215.324, "dur": 2.209,
    "args": {
      "External id": 2486014,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1381210,
    "ts": 1514657232219.481, "dur": 0.560,
    "args": {
      "External id": 2486015,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1381210,
    "ts": 1514657232221.874, "dur": 0.586,
    "args": {
      "External id": 2486016,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1381210,
    "ts": 1514657232224.314, "dur": 0.414,
    "args": {
      "External id": 2486017,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1381210,
    "ts": 1514657232226.421, "dur": 0.545,
    "args": {
      "External id": 2486018,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1381210,
    "ts": 1514657232228.570, "dur": 0.365,
    "args": {
      "External id": 2486019,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1381210,
    "ts": 1514657232230.915, "dur": 0.395,
    "args": {
      "External id": 2486020,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1381210,
    "ts": 1514657232244.617, "dur": 20.850,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657232318.177, "dur": 109.181,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657232340.760, "dur": 83.028,
    "args": {
      "External id": 2486023,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657232350.323, "dur": 69.283,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657232440.499, "dur": 1.658,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514657232516.041, "dur": 1691.512,
    "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": 1336753, "tid": 1381210, "ts": 1514657232516.041,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1381210,
    "ts": 1514657232639.792, "dur": 101.079,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657232779.597, "dur": 38.760,
    "args": {
      "External id": 2486028,"kernel_hash": "c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/5t/c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr.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": 1336753, "tid": 1381210,
    "ts": 1514657232836.361, "dur": 49.020,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657232898.043, "dur": 33.600,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657232939.625, "dur": 84.322,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657233036.752, "dur": 34.460,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657233078.964, "dur": 43.106,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657233149.368, "dur": 24.181,
    "args": {
      "External id": 2486034,"kernel_hash": "cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/mm/cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k.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": 1336753, "tid": 1381210,
    "ts": 1514657233192.506, "dur": 27.790,
    "args": {
      "External id": 2486035,"kernel_hash": "c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2", "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/4b/c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2.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": 1336753, "tid": 1381210,
    "ts": 1514657233239.813, "dur": 20.800,
    "args": {
      "External id": 2486036,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657233273.955, "dur": 17.031,
    "args": {
      "External id": 2486037,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657233302.290, "dur": 31.446,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657233337.292, "dur": 36.564,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657233402.465, "dur": 200.952,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657233483.057, "dur": 5.644,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657233491.225, "dur": 2.507,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657233639.839, "dur": 26.979,
    "args": {
      "External id": 2486043,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657233678.612, "dur": 18.752,
    "args": {
      "External id": 2486044,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657233706.291, "dur": 39.209,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657233752.628, "dur": 35.669,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657233797.614, "dur": 24.257,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657233827.662, "dur": 31.127,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657233865.426, "dur": 22.189,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657233894.959, "dur": 29.182,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657233945.176, "dur": 67.407,
    "args": {
      "External id": 2486051,"kernel_hash": "cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/as/cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf.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": 1336753, "tid": 1381210,
    "ts": 1514657234042.641, "dur": 31.019,
    "args": {
      "External id": 2486052,"kernel_hash": "cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od", "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/ny/cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od.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": 1336753, "tid": 1381210,
    "ts": 1514657234089.725, "dur": 21.827,
    "args": {
      "External id": 2486053,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657234128.533, "dur": 16.790,
    "args": {
      "External id": 2486054,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657234162.024, "dur": 17.137,
    "args": {
      "External id": 2486055,"kernel_hash": "c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/25/c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6.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": 1336753, "tid": 1381210,
    "ts": 1514657234250.113, "dur": 14.954,
    "args": {
      "External id": 2486056,"Record function id": 0, "Ev Idx": 2855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657234253.392, "dur": 10.834,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234257.265, "dur": 6.065,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234259.164, "dur": 4.043,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234269.092, "dur": 7.457,
    "args": {
      "External id": 2486060,"Record function id": 0, "Ev Idx": 2859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657234270.507, "dur": 5.563,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234271.261, "dur": 4.324,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234271.942, "dur": 3.529,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234279.990, "dur": 4.526,
    "args": {
      "External id": 2486064,"Record function id": 0, "Ev Idx": 2863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657234281.376, "dur": 2.654,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234281.830, "dur": 1.705,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234282.406, "dur": 1.044,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234287.895, "dur": 3.699,
    "args": {
      "External id": 2486068,"Record function id": 0, "Ev Idx": 2867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657234289.068, "dur": 2.116,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234289.534, "dur": 1.255,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234290.018, "dur": 0.690,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234294.956, "dur": 3.532,
    "args": {
      "External id": 2486072,"Record function id": 0, "Ev Idx": 2871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657234296.143, "dur": 1.940,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234296.588, "dur": 1.118,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234296.902, "dur": 0.734,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234301.835, "dur": 3.575,
    "args": {
      "External id": 2486076,"Record function id": 0, "Ev Idx": 2875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657234302.877, "dur": 2.088,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234303.346, "dur": 1.219,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234303.809, "dur": 0.656,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234308.822, "dur": 3.797,
    "args": {
      "External id": 2486080,"Record function id": 0, "Ev Idx": 2879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657234309.990, "dur": 2.222,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234310.459, "dur": 1.369,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234310.974, "dur": 0.745,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234315.814, "dur": 3.981,
    "args": {
      "External id": 2486084,"Record function id": 0, "Ev Idx": 2883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657234317.277, "dur": 2.097,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234317.950, "dur": 1.009,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234318.232, "dur": 0.660,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234322.917, "dur": 5.886,
    "args": {
      "External id": 2486088,"Record function id": 0, "Ev Idx": 2887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657234324.161, "dur": 4.236,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234324.816, "dur": 3.199,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234325.228, "dur": 2.684,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234332.561, "dur": 38422.686,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234333.887, "dur": 38413.683,
    "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": 1336753, "tid": 1381210, "ts": 1514657234333.887,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.19)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657234360.745, "dur": 34.967,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234403.712, "dur": 60.182,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234469.632, "dur": 38269.866,
    "args": {
      "External id": 2486096,"Record function id": 0, "Ev Idx": 2895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657234573.819, "dur": 7.563,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234591.994, "dur": 5.031,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234612.597, "dur": 37298.454,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234626.040, "dur": 37276.374,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234713.840, "dur": 13.494,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234734.141, "dur": 37128.285,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234736.696, "dur": 37124.867,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234741.027, "dur": 5.649,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657234748.507, "dur": 37109.714,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657272021.452, "dur": 9.090,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657272024.764, "dur": 5.274,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657272060.600, "dur": 364.497,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657272087.138, "dur": 333.306,
    "args": {
      "External id": 2486109,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657272097.970, "dur": 316.932,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657272442.575, "dur": 2.269,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1381210,
    "ts": 1514657272499.234, "dur": 6.797,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657272571.058, "dur": 4.089,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657272594.077, "dur": 1.056,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657272607.834, "dur": 1.186,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657272620.945, "dur": 0.896,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657272633.658, "dur": 2.997,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657272649.906, "dur": 0.905,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657272662.400, "dur": 1.750,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657272676.581, "dur": 0.775,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657272771.574, "dur": 2797.042,
    "args": {
      "External id": 2486121,"Record function id": 0, "Ev Idx": 2920
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.18)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657272789.421, "dur": 1057.734,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657272804.171, "dur": 366.445,
    "args": {
      "External id": 2486123,"Record function id": 0, "Ev Idx": 2922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1381210,
    "ts": 1514657272879.763, "dur": 5.600,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657272888.313, "dur": 1.356,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657272891.464, "dur": 0.859,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657272893.955, "dur": 0.826,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657272896.332, "dur": 0.996,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657272898.971, "dur": 0.669,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657272901.286, "dur": 2.113,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657272905.078, "dur": 1.073,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657272907.569, "dur": 2.971,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657272912.185, "dur": 0.810,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657272930.507, "dur": 204.161,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657272946.396, "dur": 183.088,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657272972.734, "dur": 53.354,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657273030.416, "dur": 68.120,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657273033.238, "dur": 64.940,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657273037.537, "dur": 6.414,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657273046.190, "dur": 51.449,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657273252.176, "dur": 586.810,
    "args": {
      "External id": 2486141,"Record function id": 0, "Ev Idx": 2940
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.17)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657273271.238, "dur": 555.465,
    "args": {
      "External id": 2486142,"Record function id": 0, "Ev Idx": 2941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657273328.622, "dur": 5.116,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657273348.887, "dur": 36.931,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657273353.988, "dur": 1.684,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657273357.806, "dur": 0.483,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657273360.080, "dur": 0.600,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657273362.511, "dur": 2.423,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657273366.586, "dur": 0.419,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657273368.989, "dur": 0.713,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657273376.465, "dur": 0.391,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657273378.575, "dur": 0.353,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657273380.996, "dur": 0.581,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657273396.070, "dur": 32.863,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657273463.680, "dur": 127.678,
    "args": {
      "External id": 2486155,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657273473.661, "dur": 3.307,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657273482.040, "dur": 9.595,
    "args": {
      "External id": 2486157,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657273486.059, "dur": 5.167,
    "args": {
      "External id": 2486158,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1381210,
    "ts": 1514657273489.664, "dur": 0.455,
    "args": {
      "External id": 2486159,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657273498.046, "dur": 28.312,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657273500.486, "dur": 2.191,
    "args": {
      "External id": 2486161,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657273504.168, "dur": 0.401,
    "args": {
      "External id": 2486162,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1381210,
    "ts": 1514657273506.227, "dur": 0.425,
    "args": {
      "External id": 2486163,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1381210,
    "ts": 1514657273508.329, "dur": 0.429,
    "args": {
      "External id": 2486164,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1381210,
    "ts": 1514657273510.926, "dur": 0.515,
    "args": {
      "External id": 2486165,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1381210,
    "ts": 1514657273513.571, "dur": 0.686,
    "args": {
      "External id": 2486166,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1381210,
    "ts": 1514657273516.214, "dur": 0.684,
    "args": {
      "External id": 2486167,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1381210,
    "ts": 1514657273518.404, "dur": 0.553,
    "args": {
      "External id": 2486168,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1381210,
    "ts": 1514657273520.804, "dur": 2.452,
    "args": {
      "External id": 2486169,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1381210,
    "ts": 1514657273559.103, "dur": 24.573,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657273638.024, "dur": 119.749,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657273665.018, "dur": 89.243,
    "args": {
      "External id": 2486172,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657273674.399, "dur": 72.008,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657273775.017, "dur": 1.825,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514657273853.897, "dur": 1694.738,
    "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": 1336753, "tid": 1381210, "ts": 1514657273853.897,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1381210,
    "ts": 1514657273955.890, "dur": 140.140,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657274137.488, "dur": 39.263,
    "args": {
      "External id": 2486177,"kernel_hash": "c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/5t/c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr.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": 1336753, "tid": 1381210,
    "ts": 1514657274197.342, "dur": 51.323,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657274258.550, "dur": 33.917,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657274299.882, "dur": 45.669,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657274353.073, "dur": 30.289,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657274393.074, "dur": 42.496,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657274461.940, "dur": 22.657,
    "args": {
      "External id": 2486183,"kernel_hash": "cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/mm/cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k.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": 1336753, "tid": 1381210,
    "ts": 1514657274506.242, "dur": 44.062,
    "args": {
      "External id": 2486184,"kernel_hash": "c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2", "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/4b/c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2.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": 1336753, "tid": 1381210,
    "ts": 1514657274577.573, "dur": 21.456,
    "args": {
      "External id": 2486185,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657274612.291, "dur": 20.708,
    "args": {
      "External id": 2486186,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657274645.375, "dur": 40.610,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657274689.474, "dur": 33.606,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657274751.885, "dur": 169.469,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657274828.715, "dur": 5.616,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657274836.407, "dur": 2.259,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657274953.360, "dur": 25.410,
    "args": {
      "External id": 2486192,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657275028.937, "dur": 19.046,
    "args": {
      "External id": 2486193,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657275059.375, "dur": 43.634,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275112.880, "dur": 38.163,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275158.953, "dur": 23.404,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275188.396, "dur": 42.202,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275243.245, "dur": 29.029,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275281.280, "dur": 33.758,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275334.239, "dur": 27.266,
    "args": {
      "External id": 2486200,"kernel_hash": "cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/as/cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf.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": 1336753, "tid": 1381210,
    "ts": 1514657275377.754, "dur": 28.194,
    "args": {
      "External id": 2486201,"kernel_hash": "cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od", "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/ny/cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od.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": 1336753, "tid": 1381210,
    "ts": 1514657275420.560, "dur": 18.721,
    "args": {
      "External id": 2486202,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657275454.028, "dur": 18.714,
    "args": {
      "External id": 2486203,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657275484.723, "dur": 16.148,
    "args": {
      "External id": 2486204,"kernel_hash": "c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/25/c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6.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": 1336753, "tid": 1381210,
    "ts": 1514657275592.057, "dur": 16.949,
    "args": {
      "External id": 2486205,"Record function id": 0, "Ev Idx": 3004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657275595.151, "dur": 12.881,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275599.201, "dur": 7.922,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275600.970, "dur": 5.985,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275613.073, "dur": 5.438,
    "args": {
      "External id": 2486209,"Record function id": 0, "Ev Idx": 3008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657275614.590, "dur": 3.474,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275615.499, "dur": 2.055,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275616.345, "dur": 1.097,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275621.992, "dur": 4.682,
    "args": {
      "External id": 2486213,"Record function id": 0, "Ev Idx": 3012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657275623.332, "dur": 2.942,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275624.050, "dur": 1.626,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275624.561, "dur": 1.021,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275630.103, "dur": 3.896,
    "args": {
      "External id": 2486217,"Record function id": 0, "Ev Idx": 3016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657275631.335, "dur": 2.246,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275631.795, "dur": 1.365,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275632.290, "dur": 0.789,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275637.339, "dur": 3.513,
    "args": {
      "External id": 2486221,"Record function id": 0, "Ev Idx": 3020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657275638.571, "dur": 1.886,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275639.012, "dur": 1.007,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275639.281, "dur": 0.664,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275644.274, "dur": 3.485,
    "args": {
      "External id": 2486225,"Record function id": 0, "Ev Idx": 3024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657275645.252, "dur": 2.074,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275645.676, "dur": 1.155,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275646.086, "dur": 0.643,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275650.954, "dur": 3.271,
    "args": {
      "External id": 2486229,"Record function id": 0, "Ev Idx": 3028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657275651.980, "dur": 1.849,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275652.439, "dur": 0.957,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275652.707, "dur": 0.615,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275657.401, "dur": 5.262,
    "args": {
      "External id": 2486233,"Record function id": 0, "Ev Idx": 3032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657275658.485, "dur": 3.770,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275658.982, "dur": 2.825,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275659.234, "dur": 2.492,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275669.055, "dur": 3.611,
    "args": {
      "External id": 2486237,"Record function id": 0, "Ev Idx": 3036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657275670.050, "dur": 2.209,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275670.632, "dur": 1.191,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275671.066, "dur": 0.661,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275676.736, "dur": 35887.669,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275678.194, "dur": 35876.836,
    "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": 1336753, "tid": 1381210, "ts": 1514657275678.194,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.18)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657275706.880, "dur": 40.059,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275755.416, "dur": 59.230,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275820.687, "dur": 35724.273,
    "args": {
      "External id": 2486245,"Record function id": 0, "Ev Idx": 3044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657275907.861, "dur": 6.383,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275923.538, "dur": 4.712,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657275942.635, "dur": 34794.242,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657276048.425, "dur": 34678.776,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657276121.364, "dur": 17.696,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657276145.530, "dur": 34540.226,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657276148.187, "dur": 34536.783,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657276152.732, "dur": 6.578,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657276161.257, "dur": 34520.961,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657310824.722, "dur": 8.480,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657310828.022, "dur": 4.854,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657310864.543, "dur": 366.658,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657310890.561, "dur": 335.713,
    "args": {
      "External id": 2486258,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657310901.081, "dur": 319.238,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657311252.494, "dur": 2.080,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1381210,
    "ts": 1514657311314.306, "dur": 6.321,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657311366.959, "dur": 1.656,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657311385.420, "dur": 0.934,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657311400.009, "dur": 0.994,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657311412.974, "dur": 0.867,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657311427.983, "dur": 0.954,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657311441.688, "dur": 0.944,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657311455.097, "dur": 1.549,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657311467.869, "dur": 0.861,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657311581.135, "dur": 2902.645,
    "args": {
      "External id": 2486270,"Record function id": 0, "Ev Idx": 3069
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.17)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657311601.917, "dur": 1131.725,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657311616.561, "dur": 445.787,
    "args": {
      "External id": 2486272,"Record function id": 0, "Ev Idx": 3071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1381210,
    "ts": 1514657311775.246, "dur": 5.349,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657311784.550, "dur": 0.872,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657311787.507, "dur": 0.702,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657311789.982, "dur": 0.939,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657311792.675, "dur": 1.173,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657311797.061, "dur": 0.789,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657311799.384, "dur": 1.635,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657311802.577, "dur": 2.699,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657311806.535, "dur": 0.851,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657311808.883, "dur": 0.625,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657311826.932, "dur": 148.724,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657311843.029, "dur": 128.066,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657311860.588, "dur": 13.026,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657311876.740, "dur": 65.062,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657311879.474, "dur": 61.968,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657311883.821, "dur": 5.603,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657311891.306, "dur": 49.369,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657312154.208, "dur": 570.941,
    "args": {
      "External id": 2486290,"Record function id": 0, "Ev Idx": 3089
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.16)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657312172.356, "dur": 540.001,
    "args": {
      "External id": 2486291,"Record function id": 0, "Ev Idx": 3090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657312228.636, "dur": 6.292,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657312250.686, "dur": 30.389,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657312255.945, "dur": 1.711,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657312259.751, "dur": 0.636,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657312262.101, "dur": 2.484,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657312266.128, "dur": 0.409,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657312268.309, "dur": 0.549,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657312270.723, "dur": 0.364,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657312272.697, "dur": 0.397,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657312274.761, "dur": 0.541,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657312276.873, "dur": 0.361,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657312290.957, "dur": 34.542,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657312358.423, "dur": 107.495,
    "args": {
      "External id": 2486304,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657312368.559, "dur": 3.645,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657312377.591, "dur": 12.495,
    "args": {
      "External id": 2486306,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657312381.798, "dur": 7.846,
    "args": {
      "External id": 2486307,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1381210,
    "ts": 1514657312385.857, "dur": 2.656,
    "args": {
      "External id": 2486308,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657312396.543, "dur": 25.034,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657312398.681, "dur": 0.459,
    "args": {
      "External id": 2486310,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657312400.782, "dur": 0.396,
    "args": {
      "External id": 2486311,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1381210,
    "ts": 1514657312402.886, "dur": 0.482,
    "args": {
      "External id": 2486312,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1381210,
    "ts": 1514657312405.314, "dur": 0.379,
    "args": {
      "External id": 2486313,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1381210,
    "ts": 1514657312407.260, "dur": 0.334,
    "args": {
      "External id": 2486314,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1381210,
    "ts": 1514657312409.270, "dur": 0.334,
    "args": {
      "External id": 2486315,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1381210,
    "ts": 1514657312411.648, "dur": 0.493,
    "args": {
      "External id": 2486316,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1381210,
    "ts": 1514657312414.034, "dur": 2.192,
    "args": {
      "External id": 2486317,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1381210,
    "ts": 1514657312418.072, "dur": 0.353,
    "args": {
      "External id": 2486318,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1381210,
    "ts": 1514657312437.479, "dur": 21.183,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657312511.606, "dur": 131.160,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657312552.461, "dur": 86.474,
    "args": {
      "External id": 2486321,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657312563.348, "dur": 71.305,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657312659.897, "dur": 1.858,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514657312741.924, "dur": 1723.155,
    "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": 1336753, "tid": 1381210, "ts": 1514657312741.924,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1381210,
    "ts": 1514657312846.025, "dur": 98.753,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657312979.997, "dur": 79.055,
    "args": {
      "External id": 2486326,"kernel_hash": "c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/5t/c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr.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": 1336753, "tid": 1381210,
    "ts": 1514657313082.241, "dur": 56.321,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657313148.317, "dur": 34.309,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657313189.797, "dur": 49.670,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657313246.996, "dur": 33.531,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657313290.150, "dur": 43.446,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657313357.249, "dur": 22.006,
    "args": {
      "External id": 2486332,"kernel_hash": "cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/mm/cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k.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": 1336753, "tid": 1381210,
    "ts": 1514657313397.178, "dur": 27.883,
    "args": {
      "External id": 2486333,"kernel_hash": "c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2", "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/4b/c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2.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": 1336753, "tid": 1381210,
    "ts": 1514657313445.182, "dur": 18.726,
    "args": {
      "External id": 2486334,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657313476.653, "dur": 17.068,
    "args": {
      "External id": 2486335,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657313505.448, "dur": 49.956,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657313560.765, "dur": 37.282,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657313628.864, "dur": 175.694,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657313709.199, "dur": 6.034,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657313717.238, "dur": 2.274,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657313837.523, "dur": 27.054,
    "args": {
      "External id": 2486341,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657313894.183, "dur": 16.762,
    "args": {
      "External id": 2486342,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657313922.049, "dur": 42.732,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657313972.053, "dur": 99.316,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314097.865, "dur": 36.403,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314140.555, "dur": 33.421,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314182.621, "dur": 25.537,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314215.430, "dur": 29.738,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314268.851, "dur": 26.392,
    "args": {
      "External id": 2486349,"kernel_hash": "cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/as/cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf.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": 1336753, "tid": 1381210,
    "ts": 1514657314312.331, "dur": 25.671,
    "args": {
      "External id": 2486350,"kernel_hash": "cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od", "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/ny/cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od.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": 1336753, "tid": 1381210,
    "ts": 1514657314352.390, "dur": 18.817,
    "args": {
      "External id": 2486351,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657314385.634, "dur": 15.696,
    "args": {
      "External id": 2486352,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657314413.974, "dur": 21.551,
    "args": {
      "External id": 2486353,"kernel_hash": "c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/25/c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6.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": 1336753, "tid": 1381210,
    "ts": 1514657314505.441, "dur": 14.513,
    "args": {
      "External id": 2486354,"Record function id": 0, "Ev Idx": 3153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657314508.594, "dur": 10.449,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314512.603, "dur": 5.511,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314514.114, "dur": 3.860,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314523.978, "dur": 21.761,
    "args": {
      "External id": 2486358,"Record function id": 0, "Ev Idx": 3157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657314525.210, "dur": 19.444,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314526.005, "dur": 17.568,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314526.914, "dur": 16.058,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314551.339, "dur": 5.538,
    "args": {
      "External id": 2486362,"Record function id": 0, "Ev Idx": 3161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657314552.752, "dur": 3.711,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314553.866, "dur": 2.107,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314554.638, "dur": 1.253,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314560.132, "dur": 4.074,
    "args": {
      "External id": 2486366,"Record function id": 0, "Ev Idx": 3165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657314561.394, "dur": 2.384,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314562.015, "dur": 1.304,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314562.348, "dur": 0.880,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314567.365, "dur": 3.989,
    "args": {
      "External id": 2486370,"Record function id": 0, "Ev Idx": 3169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657314568.672, "dur": 2.279,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314569.369, "dur": 1.162,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314569.782, "dur": 0.687,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314574.655, "dur": 4.023,
    "args": {
      "External id": 2486374,"Record function id": 0, "Ev Idx": 3173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657314575.832, "dur": 2.347,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314576.398, "dur": 1.335,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314576.916, "dur": 0.716,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314582.148, "dur": 5.585,
    "args": {
      "External id": 2486378,"Record function id": 0, "Ev Idx": 3177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657314583.276, "dur": 4.048,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314583.740, "dur": 3.165,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314584.234, "dur": 2.604,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314590.937, "dur": 3.707,
    "args": {
      "External id": 2486382,"Record function id": 0, "Ev Idx": 3181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657314592.238, "dur": 1.989,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314592.680, "dur": 1.114,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314592.942, "dur": 0.787,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314597.771, "dur": 4.062,
    "args": {
      "External id": 2486386,"Record function id": 0, "Ev Idx": 3185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657314599.000, "dur": 2.430,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314599.470, "dur": 1.532,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314600.237, "dur": 0.673,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314605.523, "dur": 36320.105,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314606.777, "dur": 36310.465,
    "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": 1336753, "tid": 1381210, "ts": 1514657314606.777,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.17)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657314638.019, "dur": 36.589,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314683.032, "dur": 62.732,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314751.933, "dur": 36158.225,
    "args": {
      "External id": 2486394,"Record function id": 0, "Ev Idx": 3193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657314838.480, "dur": 6.585,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314854.295, "dur": 8.203,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314878.155, "dur": 35249.744,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314891.297, "dur": 35227.321,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314962.694, "dur": 13.632,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657314982.628, "dur": 35094.343,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657315019.520, "dur": 35056.777,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657315024.166, "dur": 7.270,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657315033.551, "dur": 35039.352,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657350217.033, "dur": 9.015,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657350220.597, "dur": 5.107,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657350253.834, "dur": 358.874,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657350281.173, "dur": 326.446,
    "args": {
      "External id": 2486407,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657350291.777, "dur": 309.665,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657350633.113, "dur": 1.965,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1381210,
    "ts": 1514657350692.139, "dur": 8.351,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657350745.524, "dur": 1.593,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657350765.353, "dur": 1.195,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657350781.271, "dur": 0.996,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657350794.681, "dur": 3.175,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657350808.947, "dur": 0.934,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657350822.158, "dur": 0.864,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657350835.760, "dur": 1.948,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657350849.413, "dur": 2.739,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657350939.978, "dur": 2771.196,
    "args": {
      "External id": 2486419,"Record function id": 0, "Ev Idx": 3218
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.16)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657350958.160, "dur": 1062.405,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657350971.801, "dur": 353.102,
    "args": {
      "External id": 2486421,"Record function id": 0, "Ev Idx": 3220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1381210,
    "ts": 1514657351085.708, "dur": 4.872,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351094.189, "dur": 1.047,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351097.306, "dur": 1.180,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351100.091, "dur": 1.026,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351102.754, "dur": 0.852,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351105.406, "dur": 0.682,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351107.923, "dur": 3.937,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351113.236, "dur": 1.049,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351115.769, "dur": 0.683,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351117.874, "dur": 0.870,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351137.184, "dur": 154.471,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351152.876, "dur": 134.511,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351179.289, "dur": 11.786,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351194.128, "dur": 64.182,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351196.865, "dur": 61.031,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351200.984, "dur": 5.249,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351208.014, "dur": 49.227,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351409.264, "dur": 565.007,
    "args": {
      "External id": 2486439,"Record function id": 0, "Ev Idx": 3238
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.15)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657351426.975, "dur": 534.815,
    "args": {
      "External id": 2486440,"Record function id": 0, "Ev Idx": 3239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657351486.325, "dur": 4.617,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351509.210, "dur": 50.363,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351514.200, "dur": 1.860,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351518.258, "dur": 2.801,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351522.784, "dur": 0.532,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351525.326, "dur": 0.426,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351543.815, "dur": 0.801,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351547.452, "dur": 0.350,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351549.638, "dur": 0.426,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351551.568, "dur": 0.349,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351554.097, "dur": 0.482,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351571.003, "dur": 34.768,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351639.713, "dur": 104.397,
    "args": {
      "External id": 2486453,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657351650.704, "dur": 5.499,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351661.772, "dur": 9.730,
    "args": {
      "External id": 2486455,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657351665.764, "dur": 5.311,
    "args": {
      "External id": 2486456,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1381210,
    "ts": 1514657351669.451, "dur": 0.507,
    "args": {
      "External id": 2486457,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657351678.915, "dur": 26.094,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351681.342, "dur": 0.556,
    "args": {
      "External id": 2486459,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657351683.839, "dur": 0.459,
    "args": {
      "External id": 2486460,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1381210,
    "ts": 1514657351686.168, "dur": 0.450,
    "args": {
      "External id": 2486461,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1381210,
    "ts": 1514657351688.373, "dur": 0.681,
    "args": {
      "External id": 2486462,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1381210,
    "ts": 1514657351690.801, "dur": 0.566,
    "args": {
      "External id": 2486463,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1381210,
    "ts": 1514657351692.945, "dur": 0.393,
    "args": {
      "External id": 2486464,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1381210,
    "ts": 1514657351695.012, "dur": 2.133,
    "args": {
      "External id": 2486465,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1381210,
    "ts": 1514657351698.632, "dur": 0.475,
    "args": {
      "External id": 2486466,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1381210,
    "ts": 1514657351700.943, "dur": 0.386,
    "args": {
      "External id": 2486467,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1381210,
    "ts": 1514657351717.381, "dur": 19.619,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351790.808, "dur": 109.854,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351812.645, "dur": 84.565,
    "args": {
      "External id": 2486470,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657351822.057, "dur": 71.177,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657351913.898, "dur": 1.847,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514657352029.880, "dur": 1659.520,
    "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": 1336753, "tid": 1381210, "ts": 1514657352029.880,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1381210,
    "ts": 1514657352137.223, "dur": 101.701,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657352275.847, "dur": 38.663,
    "args": {
      "External id": 2486475,"kernel_hash": "c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/5t/c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr.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": 1336753, "tid": 1381210,
    "ts": 1514657352335.164, "dur": 48.387,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657352393.170, "dur": 33.256,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657352433.935, "dur": 45.693,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657352506.869, "dur": 51.466,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657352570.062, "dur": 46.789,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657352641.387, "dur": 27.315,
    "args": {
      "External id": 2486481,"kernel_hash": "cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/mm/cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k.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": 1336753, "tid": 1381210,
    "ts": 1514657352686.633, "dur": 28.587,
    "args": {
      "External id": 2486482,"kernel_hash": "c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2", "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/4b/c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2.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": 1336753, "tid": 1381210,
    "ts": 1514657352739.604, "dur": 19.161,
    "args": {
      "External id": 2486483,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657352772.902, "dur": 17.566,
    "args": {
      "External id": 2486484,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657352802.584, "dur": 30.103,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657352835.992, "dur": 32.332,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657352896.227, "dur": 229.921,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657352977.039, "dur": 44.876,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353025.259, "dur": 4.287,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353161.252, "dur": 25.151,
    "args": {
      "External id": 2486490,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657353197.585, "dur": 14.917,
    "args": {
      "External id": 2486491,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657353220.814, "dur": 39.504,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353266.440, "dur": 33.760,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353306.496, "dur": 22.301,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353335.080, "dur": 28.245,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353368.620, "dur": 20.823,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353394.657, "dur": 35.963,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353454.310, "dur": 27.523,
    "args": {
      "External id": 2486498,"kernel_hash": "cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/as/cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf.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": 1336753, "tid": 1381210,
    "ts": 1514657353498.305, "dur": 25.101,
    "args": {
      "External id": 2486499,"kernel_hash": "cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od", "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/ny/cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od.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": 1336753, "tid": 1381210,
    "ts": 1514657353555.244, "dur": 27.089,
    "args": {
      "External id": 2486500,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657353606.549, "dur": 21.283,
    "args": {
      "External id": 2486501,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657353644.781, "dur": 17.398,
    "args": {
      "External id": 2486502,"kernel_hash": "c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/25/c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6.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": 1336753, "tid": 1381210,
    "ts": 1514657353732.686, "dur": 14.749,
    "args": {
      "External id": 2486503,"Record function id": 0, "Ev Idx": 3302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657353735.768, "dur": 10.723,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353739.882, "dur": 5.833,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353741.729, "dur": 3.889,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353751.870, "dur": 8.738,
    "args": {
      "External id": 2486507,"Record function id": 0, "Ev Idx": 3306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657353756.736, "dur": 3.308,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353757.883, "dur": 1.656,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353758.549, "dur": 0.899,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353764.618, "dur": 4.376,
    "args": {
      "External id": 2486511,"Record function id": 0, "Ev Idx": 3310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657353765.714, "dur": 2.856,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353766.495, "dur": 1.660,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353767.072, "dur": 1.005,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353772.665, "dur": 6.365,
    "args": {
      "External id": 2486515,"Record function id": 0, "Ev Idx": 3314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657353773.872, "dur": 4.756,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353774.746, "dur": 3.485,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353775.476, "dur": 2.659,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353782.842, "dur": 4.988,
    "args": {
      "External id": 2486519,"Record function id": 0, "Ev Idx": 3318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657353784.450, "dur": 2.938,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353785.295, "dur": 1.573,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353785.986, "dur": 0.815,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353791.289, "dur": 4.065,
    "args": {
      "External id": 2486523,"Record function id": 0, "Ev Idx": 3322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657353792.638, "dur": 2.292,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353793.380, "dur": 1.127,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353793.852, "dur": 0.569,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353798.757, "dur": 4.029,
    "args": {
      "External id": 2486527,"Record function id": 0, "Ev Idx": 3326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657353799.847, "dur": 2.512,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353800.486, "dur": 1.486,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353800.971, "dur": 0.923,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353806.182, "dur": 3.912,
    "args": {
      "External id": 2486531,"Record function id": 0, "Ev Idx": 3330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657353807.381, "dur": 2.311,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353808.014, "dur": 1.266,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353808.572, "dur": 0.629,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353813.643, "dur": 4.350,
    "args": {
      "External id": 2486535,"Record function id": 0, "Ev Idx": 3334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657353814.880, "dur": 2.692,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353815.672, "dur": 1.518,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353816.508, "dur": 0.618,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353821.978, "dur": 36452.847,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353823.397, "dur": 36443.136,
    "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": 1336753, "tid": 1381210, "ts": 1514657353823.397,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.16)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657353848.705, "dur": 32.980,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353888.700, "dur": 61.187,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657353955.512, "dur": 36303.698,
    "args": {
      "External id": 2486543,"Record function id": 0, "Ev Idx": 3342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657354080.280, "dur": 7.519,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657354098.086, "dur": 4.937,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657354119.951, "dur": 35311.638,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657354133.664, "dur": 35289.263,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657354202.405, "dur": 14.900,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657354223.689, "dur": 35158.340,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657354229.715, "dur": 35151.688,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657354234.318, "dur": 5.528,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657354241.709, "dur": 35136.505,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657389514.835, "dur": 8.075,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657389517.790, "dur": 4.780,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657389562.561, "dur": 357.361,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657389592.638, "dur": 322.059,
    "args": {
      "External id": 2486556,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657389603.935, "dur": 305.136,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657389949.812, "dur": 2.374,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1381210,
    "ts": 1514657390042.067, "dur": 6.854,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390094.474, "dur": 1.159,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390113.991, "dur": 1.035,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390130.283, "dur": 0.826,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390143.774, "dur": 1.080,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390157.093, "dur": 1.060,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390170.516, "dur": 1.029,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390186.439, "dur": 1.723,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390200.202, "dur": 0.732,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390291.519, "dur": 2744.937,
    "args": {
      "External id": 2486568,"Record function id": 0, "Ev Idx": 3367
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.15)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657390310.074, "dur": 997.346,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390324.254, "dur": 330.025,
    "args": {
      "External id": 2486570,"Record function id": 0, "Ev Idx": 3369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1381210,
    "ts": 1514657390402.558, "dur": 4.213,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390410.063, "dur": 1.345,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390413.551, "dur": 0.972,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390416.225, "dur": 3.106,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390420.749, "dur": 0.974,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390423.791, "dur": 1.048,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390426.643, "dur": 1.722,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390429.744, "dur": 1.334,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390432.597, "dur": 1.163,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390435.431, "dur": 1.149,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390453.883, "dur": 166.631,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390469.427, "dur": 146.289,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390490.400, "dur": 11.540,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390505.180, "dur": 81.305,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390508.231, "dur": 77.868,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390512.482, "dur": 6.975,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390521.521, "dur": 63.667,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390735.190, "dur": 563.921,
    "args": {
      "External id": 2486588,"Record function id": 0, "Ev Idx": 3387
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.14)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657390752.478, "dur": 534.699,
    "args": {
      "External id": 2486589,"Record function id": 0, "Ev Idx": 3388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657390806.812, "dur": 5.297,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390826.785, "dur": 31.587,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390831.969, "dur": 1.629,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390835.815, "dur": 0.522,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390838.226, "dur": 0.593,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390840.717, "dur": 0.430,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390843.144, "dur": 0.393,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390845.523, "dur": 0.412,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390847.703, "dur": 2.766,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390852.564, "dur": 0.371,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390854.283, "dur": 0.461,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390867.820, "dur": 31.317,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390929.873, "dur": 143.562,
    "args": {
      "External id": 2486602,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657390939.833, "dur": 3.279,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390948.370, "dur": 9.700,
    "args": {
      "External id": 2486604,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657390952.612, "dur": 5.021,
    "args": {
      "External id": 2486605,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1381210,
    "ts": 1514657390956.241, "dur": 0.375,
    "args": {
      "External id": 2486606,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657390964.792, "dur": 63.433,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657390967.097, "dur": 0.480,
    "args": {
      "External id": 2486608,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657390969.651, "dur": 0.441,
    "args": {
      "External id": 2486609,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1381210,
    "ts": 1514657390971.690, "dur": 0.484,
    "args": {
      "External id": 2486610,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1381210,
    "ts": 1514657390974.031, "dur": 2.503,
    "args": {
      "External id": 2486611,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1381210,
    "ts": 1514657390978.004, "dur": 0.416,
    "args": {
      "External id": 2486612,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1381210,
    "ts": 1514657390980.316, "dur": 0.373,
    "args": {
      "External id": 2486613,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1381210,
    "ts": 1514657391018.882, "dur": 0.688,
    "args": {
      "External id": 2486614,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1381210,
    "ts": 1514657391021.894, "dur": 0.329,
    "args": {
      "External id": 2486615,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1381210,
    "ts": 1514657391023.940, "dur": 0.376,
    "args": {
      "External id": 2486616,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1381210,
    "ts": 1514657391040.527, "dur": 25.172,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657391117.926, "dur": 107.674,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657391139.487, "dur": 82.618,
    "args": {
      "External id": 2486619,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657391149.024, "dur": 68.671,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657391238.264, "dur": 1.888,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514657391314.767, "dur": 1664.139,
    "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": 1336753, "tid": 1381210, "ts": 1514657391314.767,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1381210,
    "ts": 1514657391416.249, "dur": 97.370,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657391570.261, "dur": 47.191,
    "args": {
      "External id": 2486624,"kernel_hash": "c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/5t/c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr.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": 1336753, "tid": 1381210,
    "ts": 1514657391636.434, "dur": 54.191,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657391700.411, "dur": 33.628,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657391743.227, "dur": 45.793,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657391796.611, "dur": 29.031,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657391832.983, "dur": 41.488,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657391896.916, "dur": 22.859,
    "args": {
      "External id": 2486630,"kernel_hash": "cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/mm/cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k.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": 1336753, "tid": 1381210,
    "ts": 1514657391940.218, "dur": 30.688,
    "args": {
      "External id": 2486631,"kernel_hash": "c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2", "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/4b/c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2.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": 1336753, "tid": 1381210,
    "ts": 1514657392045.440, "dur": 25.050,
    "args": {
      "External id": 2486632,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657392086.876, "dur": 16.767,
    "args": {
      "External id": 2486633,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657392113.361, "dur": 37.118,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657392154.132, "dur": 33.552,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657392233.197, "dur": 169.823,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657392310.328, "dur": 5.408,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657392317.954, "dur": 2.405,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657392432.896, "dur": 25.165,
    "args": {
      "External id": 2486639,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657392469.735, "dur": 15.601,
    "args": {
      "External id": 2486640,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657392493.522, "dur": 52.119,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657392555.796, "dur": 44.407,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657392610.097, "dur": 25.825,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657392641.405, "dur": 29.903,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657392677.924, "dur": 21.716,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657392706.930, "dur": 45.908,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657392777.985, "dur": 25.396,
    "args": {
      "External id": 2486647,"kernel_hash": "cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/as/cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf.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": 1336753, "tid": 1381210,
    "ts": 1514657392820.569, "dur": 28.656,
    "args": {
      "External id": 2486648,"kernel_hash": "cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od", "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/ny/cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od.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": 1336753, "tid": 1381210,
    "ts": 1514657392863.460, "dur": 17.548,
    "args": {
      "External id": 2486649,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657392900.808, "dur": 16.935,
    "args": {
      "External id": 2486650,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657392930.124, "dur": 16.933,
    "args": {
      "External id": 2486651,"kernel_hash": "c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/25/c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6.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": 1336753, "tid": 1381210,
    "ts": 1514657393060.767, "dur": 15.617,
    "args": {
      "External id": 2486652,"Record function id": 0, "Ev Idx": 3451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657393064.476, "dur": 11.023,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393068.348, "dur": 6.306,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393070.284, "dur": 4.237,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393080.355, "dur": 9.404,
    "args": {
      "External id": 2486656,"Record function id": 0, "Ev Idx": 3455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657393085.220, "dur": 4.133,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393086.184, "dur": 2.669,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393087.496, "dur": 1.259,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393092.945, "dur": 6.347,
    "args": {
      "External id": 2486660,"Record function id": 0, "Ev Idx": 3459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657393094.293, "dur": 4.597,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393094.812, "dur": 3.587,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393095.363, "dur": 2.968,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393102.519, "dur": 4.639,
    "args": {
      "External id": 2486664,"Record function id": 0, "Ev Idx": 3463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657393103.854, "dur": 2.873,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393104.659, "dur": 1.638,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393105.396, "dur": 0.827,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393110.360, "dur": 3.766,
    "args": {
      "External id": 2486668,"Record function id": 0, "Ev Idx": 3467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657393111.548, "dur": 2.177,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393112.206, "dur": 1.058,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393112.581, "dur": 0.619,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393117.309, "dur": 4.274,
    "args": {
      "External id": 2486672,"Record function id": 0, "Ev Idx": 3471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657393118.877, "dur": 2.295,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393119.519, "dur": 1.198,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393119.950, "dur": 0.645,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393124.845, "dur": 3.865,
    "args": {
      "External id": 2486676,"Record function id": 0, "Ev Idx": 3475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657393126.140, "dur": 2.150,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393126.675, "dur": 1.195,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393127.155, "dur": 0.629,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393131.909, "dur": 3.711,
    "args": {
      "External id": 2486680,"Record function id": 0, "Ev Idx": 3479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657393133.145, "dur": 2.047,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393133.716, "dur": 1.031,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393133.998, "dur": 0.642,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393138.738, "dur": 7.252,
    "args": {
      "External id": 2486684,"Record function id": 0, "Ev Idx": 3483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657393143.293, "dur": 2.287,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393143.741, "dur": 1.417,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393144.420, "dur": 0.621,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393149.730, "dur": 36682.305,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393151.451, "dur": 36672.911,
    "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": 1336753, "tid": 1381210, "ts": 1514657393151.451,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.15)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657393181.010, "dur": 39.564,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393228.540, "dur": 63.208,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393297.697, "dur": 36519.122,
    "args": {
      "External id": 2486692,"Record function id": 0, "Ev Idx": 3491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657393384.099, "dur": 6.715,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393399.707, "dur": 6.689,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393421.418, "dur": 35630.424,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393434.609, "dur": 35607.993,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393501.566, "dur": 12.960,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393521.003, "dur": 35452.794,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393523.876, "dur": 35449.138,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393544.673, "dur": 7.680,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657393554.989, "dur": 35414.967,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657429138.432, "dur": 8.696,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657429141.854, "dur": 4.926,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657429178.790, "dur": 328.438,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657429205.087, "dur": 297.737,
    "args": {
      "External id": 2486705,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657429215.520, "dur": 282.024,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657429524.428, "dur": 2.278,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1381210,
    "ts": 1514657429601.769, "dur": 6.639,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657429651.755, "dur": 1.394,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657429670.725, "dur": 3.044,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657429687.715, "dur": 0.968,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657429701.107, "dur": 0.834,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657429714.207, "dur": 0.887,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657429727.630, "dur": 2.941,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657429743.497, "dur": 1.989,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657429757.599, "dur": 0.729,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657429848.124, "dur": 2888.917,
    "args": {
      "External id": 2486717,"Record function id": 0, "Ev Idx": 3516
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.14)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657429865.836, "dur": 1020.109,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657429881.223, "dur": 356.198,
    "args": {
      "External id": 2486719,"Record function id": 0, "Ev Idx": 3518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1381210,
    "ts": 1514657429958.380, "dur": 3.864,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657429965.131, "dur": 0.966,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657429967.905, "dur": 3.035,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657429972.433, "dur": 1.031,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657429975.104, "dur": 1.284,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657429978.151, "dur": 1.109,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657429980.951, "dur": 36.503,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657430021.539, "dur": 0.974,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657430024.440, "dur": 0.873,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657430026.935, "dur": 1.098,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657430046.322, "dur": 157.151,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657430061.991, "dur": 137.170,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657430086.793, "dur": 14.160,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657430104.035, "dur": 65.672,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657430106.652, "dur": 62.736,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657430110.724, "dur": 5.832,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657430118.524, "dur": 50.304,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657430316.537, "dur": 560.615,
    "args": {
      "External id": 2486737,"Record function id": 0, "Ev Idx": 3536
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.13)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657430334.641, "dur": 530.431,
    "args": {
      "External id": 2486738,"Record function id": 0, "Ev Idx": 3537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657430390.033, "dur": 4.736,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657430410.213, "dur": 31.308,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657430415.505, "dur": 1.879,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657430419.561, "dur": 0.629,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657430422.214, "dur": 0.401,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657430424.000, "dur": 0.465,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657430426.147, "dur": 0.332,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657430428.392, "dur": 2.273,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657430432.528, "dur": 0.561,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657430434.746, "dur": 0.385,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657430437.256, "dur": 0.411,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657430451.888, "dur": 31.519,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657430513.469, "dur": 129.468,
    "args": {
      "External id": 2486751,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657430522.575, "dur": 2.724,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657430553.323, "dur": 11.562,
    "args": {
      "External id": 2486753,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657430557.641, "dur": 6.801,
    "args": {
      "External id": 2486754,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1381210,
    "ts": 1514657430561.905, "dur": 0.784,
    "args": {
      "External id": 2486755,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657430575.265, "dur": 24.825,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657430577.399, "dur": 0.466,
    "args": {
      "External id": 2486757,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657430579.619, "dur": 0.455,
    "args": {
      "External id": 2486758,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1381210,
    "ts": 1514657430582.001, "dur": 2.282,
    "args": {
      "External id": 2486759,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1381210,
    "ts": 1514657430585.707, "dur": 0.554,
    "args": {
      "External id": 2486760,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1381210,
    "ts": 1514657430587.859, "dur": 0.383,
    "args": {
      "External id": 2486761,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1381210,
    "ts": 1514657430589.970, "dur": 0.367,
    "args": {
      "External id": 2486762,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1381210,
    "ts": 1514657430591.990, "dur": 0.484,
    "args": {
      "External id": 2486763,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1381210,
    "ts": 1514657430593.946, "dur": 0.391,
    "args": {
      "External id": 2486764,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1381210,
    "ts": 1514657430595.920, "dur": 0.455,
    "args": {
      "External id": 2486765,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1381210,
    "ts": 1514657430611.952, "dur": 23.713,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657430689.822, "dur": 109.049,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657430711.935, "dur": 83.424,
    "args": {
      "External id": 2486768,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657430721.480, "dur": 69.269,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657430812.446, "dur": 1.690,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514657430893.592, "dur": 1825.241,
    "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": 1336753, "tid": 1381210, "ts": 1514657430893.592,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1381210,
    "ts": 1514657431041.339, "dur": 102.138,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657431181.079, "dur": 39.864,
    "args": {
      "External id": 2486773,"kernel_hash": "c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/5t/c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr.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": 1336753, "tid": 1381210,
    "ts": 1514657431237.488, "dur": 47.429,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657431295.964, "dur": 32.674,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657431335.501, "dur": 45.494,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657431388.531, "dur": 26.882,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657431421.544, "dur": 41.825,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657431486.512, "dur": 21.337,
    "args": {
      "External id": 2486779,"kernel_hash": "cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/mm/cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k.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": 1336753, "tid": 1381210,
    "ts": 1514657431524.060, "dur": 47.776,
    "args": {
      "External id": 2486780,"kernel_hash": "c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2", "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/4b/c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2.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": 1336753, "tid": 1381210,
    "ts": 1514657431594.595, "dur": 19.230,
    "args": {
      "External id": 2486781,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657431627.996, "dur": 16.107,
    "args": {
      "External id": 2486782,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657431652.645, "dur": 34.395,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657431690.332, "dur": 36.989,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657431828.752, "dur": 222.624,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657431907.725, "dur": 5.677,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657431915.339, "dur": 1.743,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432089.802, "dur": 34.586,
    "args": {
      "External id": 2486788,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657432136.013, "dur": 15.920,
    "args": {
      "External id": 2486789,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657432166.304, "dur": 52.904,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432225.698, "dur": 43.014,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432277.971, "dur": 26.693,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432310.674, "dur": 49.119,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432372.725, "dur": 26.420,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432407.264, "dur": 35.264,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432463.618, "dur": 27.076,
    "args": {
      "External id": 2486796,"kernel_hash": "cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/as/cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf.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": 1336753, "tid": 1381210,
    "ts": 1514657432556.131, "dur": 29.461,
    "args": {
      "External id": 2486797,"kernel_hash": "cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od", "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/ny/cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od.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": 1336753, "tid": 1381210,
    "ts": 1514657432604.457, "dur": 19.835,
    "args": {
      "External id": 2486798,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657432642.707, "dur": 16.881,
    "args": {
      "External id": 2486799,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657432672.050, "dur": 17.908,
    "args": {
      "External id": 2486800,"kernel_hash": "c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/25/c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6.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": 1336753, "tid": 1381210,
    "ts": 1514657432760.372, "dur": 14.919,
    "args": {
      "External id": 2486801,"Record function id": 0, "Ev Idx": 3600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657432763.778, "dur": 10.621,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432767.654, "dur": 6.009,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432769.240, "dur": 4.283,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432779.204, "dur": 7.820,
    "args": {
      "External id": 2486805,"Record function id": 0, "Ev Idx": 3604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657432780.609, "dur": 5.987,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432781.453, "dur": 4.719,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432782.078, "dur": 3.994,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432790.178, "dur": 4.514,
    "args": {
      "External id": 2486809,"Record function id": 0, "Ev Idx": 3608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657432791.922, "dur": 2.345,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432792.479, "dur": 1.365,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432792.815, "dur": 0.927,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432797.868, "dur": 4.188,
    "args": {
      "External id": 2486813,"Record function id": 0, "Ev Idx": 3612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657432799.260, "dur": 2.403,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432799.898, "dur": 1.385,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432800.342, "dur": 0.863,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432805.131, "dur": 3.707,
    "args": {
      "External id": 2486817,"Record function id": 0, "Ev Idx": 3616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657432806.242, "dur": 2.174,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432806.703, "dur": 1.338,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432807.149, "dur": 0.817,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432812.167, "dur": 4.308,
    "args": {
      "External id": 2486821,"Record function id": 0, "Ev Idx": 3620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657432813.410, "dur": 2.660,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432814.304, "dur": 1.358,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432814.729, "dur": 0.830,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432819.834, "dur": 3.622,
    "args": {
      "External id": 2486825,"Record function id": 0, "Ev Idx": 3624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657432821.034, "dur": 2.024,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432821.671, "dur": 0.987,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432821.944, "dur": 0.622,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432826.597, "dur": 3.437,
    "args": {
      "External id": 2486829,"Record function id": 0, "Ev Idx": 3628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657432827.694, "dur": 1.917,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432828.139, "dur": 1.068,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432828.391, "dur": 0.749,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432833.055, "dur": 5.846,
    "args": {
      "External id": 2486833,"Record function id": 0, "Ev Idx": 3632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657432834.352, "dur": 4.105,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432834.788, "dur": 3.294,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432835.223, "dur": 2.766,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432842.790, "dur": 35357.090,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432844.115, "dur": 35347.524,
    "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": 1336753, "tid": 1381210, "ts": 1514657432844.115,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.14)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657432874.442, "dur": 36.518,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657432917.977, "dur": 62.238,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657433020.901, "dur": 35162.513,
    "args": {
      "External id": 2486841,"Record function id": 0, "Ev Idx": 3640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657433111.017, "dur": 7.149,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657433128.343, "dur": 5.033,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657433148.754, "dur": 34199.202,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657433161.988, "dur": 34177.278,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657433223.127, "dur": 13.238,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657433243.177, "dur": 34057.016,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657433245.750, "dur": 34053.540,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657433249.900, "dur": 5.446,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657433257.175, "dur": 34038.820,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657467434.550, "dur": 8.883,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657467438.063, "dur": 4.962,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657467474.100, "dur": 365.827,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657467499.938, "dur": 335.036,
    "args": {
      "External id": 2486854,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657467510.206, "dur": 318.655,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657467859.987, "dur": 2.358,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1381210,
    "ts": 1514657467921.438, "dur": 6.862,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657467970.805, "dur": 3.552,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468038.420, "dur": 1.878,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468054.527, "dur": 1.086,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468067.383, "dur": 0.957,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468079.862, "dur": 2.925,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468094.112, "dur": 0.832,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468107.013, "dur": 1.941,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468119.382, "dur": 0.831,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468215.828, "dur": 2737.599,
    "args": {
      "External id": 2486866,"Record function id": 0, "Ev Idx": 3665
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.13)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657468238.080, "dur": 1027.981,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468252.186, "dur": 345.584,
    "args": {
      "External id": 2486868,"Record function id": 0, "Ev Idx": 3667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1381210,
    "ts": 1514657468327.435, "dur": 6.328,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468336.777, "dur": 0.939,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468339.540, "dur": 0.699,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468342.119, "dur": 0.722,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468344.266, "dur": 1.008,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468346.982, "dur": 0.890,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468349.543, "dur": 1.850,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468353.087, "dur": 0.862,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468355.528, "dur": 2.657,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468359.661, "dur": 0.814,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468379.068, "dur": 183.610,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468394.313, "dur": 162.700,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468431.965, "dur": 12.097,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468447.177, "dur": 63.990,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468449.981, "dur": 60.767,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468453.981, "dur": 5.428,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468461.155, "dur": 48.934,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468680.632, "dur": 577.300,
    "args": {
      "External id": 2486886,"Record function id": 0, "Ev Idx": 3685
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.12)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657468697.586, "dur": 547.681,
    "args": {
      "External id": 2486887,"Record function id": 0, "Ev Idx": 3686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657468754.866, "dur": 6.464,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468777.281, "dur": 30.139,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468782.042, "dur": 1.408,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468785.787, "dur": 0.409,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468788.107, "dur": 0.626,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468790.533, "dur": 2.513,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468794.592, "dur": 0.475,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468796.597, "dur": 0.553,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468798.752, "dur": 0.357,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468801.028, "dur": 0.394,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468802.893, "dur": 0.365,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468817.765, "dur": 33.230,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468881.853, "dur": 98.100,
    "args": {
      "External id": 2486900,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657468891.453, "dur": 3.482,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468900.277, "dur": 9.882,
    "args": {
      "External id": 2486902,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657468904.332, "dur": 5.435,
    "args": {
      "External id": 2486903,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1381210,
    "ts": 1514657468907.926, "dur": 0.813,
    "args": {
      "External id": 2486904,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657468917.210, "dur": 26.535,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657468919.273, "dur": 2.816,
    "args": {
      "External id": 2486906,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657468923.589, "dur": 0.342,
    "args": {
      "External id": 2486907,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1381210,
    "ts": 1514657468925.733, "dur": 0.559,
    "args": {
      "External id": 2486908,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1381210,
    "ts": 1514657468928.014, "dur": 0.431,
    "args": {
      "External id": 2486909,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1381210,
    "ts": 1514657468929.973, "dur": 0.379,
    "args": {
      "External id": 2486910,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1381210,
    "ts": 1514657468931.819, "dur": 0.588,
    "args": {
      "External id": 2486911,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1381210,
    "ts": 1514657468934.146, "dur": 0.341,
    "args": {
      "External id": 2486912,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1381210,
    "ts": 1514657468936.006, "dur": 0.403,
    "args": {
      "External id": 2486913,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1381210,
    "ts": 1514657468938.320, "dur": 2.185,
    "args": {
      "External id": 2486914,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1381210,
    "ts": 1514657468952.732, "dur": 20.708,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657469062.706, "dur": 117.964,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657469086.015, "dur": 90.768,
    "args": {
      "External id": 2486917,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657469095.846, "dur": 76.502,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657469194.103, "dur": 1.873,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514657469273.136, "dur": 1659.569,
    "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": 1336753, "tid": 1381210, "ts": 1514657469273.136,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1381210,
    "ts": 1514657469376.406, "dur": 98.100,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657469512.764, "dur": 60.797,
    "args": {
      "External id": 2486922,"kernel_hash": "c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/5t/c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr.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": 1336753, "tid": 1381210,
    "ts": 1514657469595.471, "dur": 53.142,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657469657.534, "dur": 32.821,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657469697.798, "dur": 44.834,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657469749.813, "dur": 28.592,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657469787.622, "dur": 46.646,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657469858.265, "dur": 25.327,
    "args": {
      "External id": 2486928,"kernel_hash": "cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/mm/cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k.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": 1336753, "tid": 1381210,
    "ts": 1514657469901.330, "dur": 28.194,
    "args": {
      "External id": 2486929,"kernel_hash": "c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2", "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/4b/c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2.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": 1336753, "tid": 1381210,
    "ts": 1514657469947.704, "dur": 19.447,
    "args": {
      "External id": 2486930,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657469980.325, "dur": 56.781,
    "args": {
      "External id": 2486931,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657470051.358, "dur": 35.885,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657470091.050, "dur": 35.051,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657470152.694, "dur": 168.184,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657470228.964, "dur": 6.217,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657470237.247, "dur": 2.185,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657470352.271, "dur": 25.216,
    "args": {
      "External id": 2486937,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657470409.393, "dur": 16.262,
    "args": {
      "External id": 2486938,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657470434.226, "dur": 37.907,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657470480.491, "dur": 36.050,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657470526.535, "dur": 43.144,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657470577.764, "dur": 30.725,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657470615.223, "dur": 36.688,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657470666.820, "dur": 42.010,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657470729.058, "dur": 29.472,
    "args": {
      "External id": 2486945,"kernel_hash": "cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/as/cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf.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": 1336753, "tid": 1381210,
    "ts": 1514657470775.719, "dur": 24.658,
    "args": {
      "External id": 2486946,"kernel_hash": "cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od", "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/ny/cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od.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": 1336753, "tid": 1381210,
    "ts": 1514657470817.133, "dur": 17.934,
    "args": {
      "External id": 2486947,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657470853.072, "dur": 19.405,
    "args": {
      "External id": 2486948,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657470886.197, "dur": 16.553,
    "args": {
      "External id": 2486949,"kernel_hash": "c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/25/c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6.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": 1336753, "tid": 1381210,
    "ts": 1514657470975.405, "dur": 55.844,
    "args": {
      "External id": 2486950,"Record function id": 0, "Ev Idx": 3749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657470978.446, "dur": 51.207,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471019.900, "dur": 8.327,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471021.767, "dur": 6.145,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471037.696, "dur": 8.637,
    "args": {
      "External id": 2486954,"Record function id": 0, "Ev Idx": 3753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657471042.707, "dur": 3.196,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471043.509, "dur": 1.820,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471044.273, "dur": 0.963,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471049.670, "dur": 4.446,
    "args": {
      "External id": 2486958,"Record function id": 0, "Ev Idx": 3757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657471050.911, "dur": 2.820,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471051.468, "dur": 1.819,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471051.966, "dur": 1.231,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471057.411, "dur": 3.602,
    "args": {
      "External id": 2486962,"Record function id": 0, "Ev Idx": 3761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657471058.570, "dur": 2.048,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471059.102, "dur": 1.071,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471059.417, "dur": 0.683,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471064.133, "dur": 3.547,
    "args": {
      "External id": 2486966,"Record function id": 0, "Ev Idx": 3765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657471065.248, "dur": 2.023,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471065.798, "dur": 1.002,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471066.070, "dur": 0.661,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471070.900, "dur": 3.656,
    "args": {
      "External id": 2486970,"Record function id": 0, "Ev Idx": 3769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657471072.022, "dur": 2.153,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471072.477, "dur": 1.257,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471072.991, "dur": 0.640,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471077.919, "dur": 3.715,
    "args": {
      "External id": 2486974,"Record function id": 0, "Ev Idx": 3773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657471079.146, "dur": 2.098,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471079.608, "dur": 1.209,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471080.077, "dur": 0.643,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471084.794, "dur": 5.767,
    "args": {
      "External id": 2486978,"Record function id": 0, "Ev Idx": 3777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657471085.967, "dur": 4.210,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471086.418, "dur": 3.350,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471086.675, "dur": 3.005,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471093.637, "dur": 7.482,
    "args": {
      "External id": 2486982,"Record function id": 0, "Ev Idx": 3781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657471094.728, "dur": 5.979,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471095.234, "dur": 5.059,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471095.930, "dur": 4.268,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471104.638, "dur": 36346.717,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471106.046, "dur": 36337.671,
    "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": 1336753, "tid": 1381210, "ts": 1514657471106.046,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.13)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657471136.336, "dur": 36.519,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471180.936, "dur": 59.369,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471246.167, "dur": 36190.287,
    "args": {
      "External id": 2486990,"Record function id": 0, "Ev Idx": 3789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657471331.700, "dur": 7.225,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471347.901, "dur": 4.775,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471367.121, "dur": 35262.069,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471380.187, "dur": 35240.057,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471439.600, "dur": 13.322,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471459.026, "dur": 35120.109,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471461.591, "dur": 35116.837,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471465.590, "dur": 5.952,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657471473.408, "dur": 35101.854,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657506715.457, "dur": 8.233,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657506718.475, "dur": 4.823,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657506753.604, "dur": 388.127,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657506778.898, "dur": 357.521,
    "args": {
      "External id": 2487003,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657506789.529, "dur": 340.867,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657507162.609, "dur": 2.489,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1381210,
    "ts": 1514657507226.120, "dur": 6.541,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657507278.406, "dur": 1.342,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657507297.199, "dur": 0.829,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657507311.069, "dur": 1.048,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657507323.334, "dur": 0.922,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657507336.758, "dur": 0.887,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657507349.817, "dur": 0.985,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657507362.756, "dur": 1.896,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657507375.199, "dur": 0.877,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657507465.576, "dur": 2688.534,
    "args": {
      "External id": 2487015,"Record function id": 0, "Ev Idx": 3814
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.12)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657507483.823, "dur": 997.606,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657507497.141, "dur": 325.746,
    "args": {
      "External id": 2487017,"Record function id": 0, "Ev Idx": 3816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1381210,
    "ts": 1514657507589.844, "dur": 4.544,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657507598.026, "dur": 1.239,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657507601.037, "dur": 0.828,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657507603.508, "dur": 1.105,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657507606.303, "dur": 0.890,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657507608.730, "dur": 0.623,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657507610.750, "dur": 1.859,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657507613.913, "dur": 2.844,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657507618.176, "dur": 0.793,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657507620.609, "dur": 0.793,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657507640.166, "dur": 149.711,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657507656.013, "dur": 129.639,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657507678.776, "dur": 11.469,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657507693.406, "dur": 63.305,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657507696.024, "dur": 60.353,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657507700.200, "dur": 5.335,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657507707.444, "dur": 48.336,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657507905.293, "dur": 568.458,
    "args": {
      "External id": 2487035,"Record function id": 0, "Ev Idx": 3834
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.11)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657507921.940, "dur": 539.001,
    "args": {
      "External id": 2487036,"Record function id": 0, "Ev Idx": 3835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657507976.518, "dur": 4.353,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657508033.300, "dur": 30.621,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657508038.181, "dur": 1.633,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657508042.081, "dur": 0.612,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657508044.800, "dur": 2.554,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657508048.923, "dur": 0.328,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657508050.830, "dur": 0.399,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657508052.663, "dur": 0.602,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657508054.871, "dur": 0.382,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657508057.209, "dur": 0.436,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657508059.092, "dur": 0.455,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657508073.554, "dur": 36.896,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657508143.936, "dur": 100.638,
    "args": {
      "External id": 2487049,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657508154.447, "dur": 4.101,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657508163.874, "dur": 11.258,
    "args": {
      "External id": 2487051,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657508167.765, "dur": 6.974,
    "args": {
      "External id": 2487052,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1381210,
    "ts": 1514657508171.232, "dur": 2.408,
    "args": {
      "External id": 2487053,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657508182.044, "dur": 24.263,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657508184.304, "dur": 0.467,
    "args": {
      "External id": 2487055,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657508186.409, "dur": 0.405,
    "args": {
      "External id": 2487056,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1381210,
    "ts": 1514657508188.582, "dur": 0.427,
    "args": {
      "External id": 2487057,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1381210,
    "ts": 1514657508190.669, "dur": 0.522,
    "args": {
      "External id": 2487058,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1381210,
    "ts": 1514657508192.624, "dur": 0.431,
    "args": {
      "External id": 2487059,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1381210,
    "ts": 1514657508194.603, "dur": 0.414,
    "args": {
      "External id": 2487060,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1381210,
    "ts": 1514657508196.541, "dur": 0.417,
    "args": {
      "External id": 2487061,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1381210,
    "ts": 1514657508198.438, "dur": 2.467,
    "args": {
      "External id": 2487062,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1381210,
    "ts": 1514657508202.544, "dur": 0.406,
    "args": {
      "External id": 2487063,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1381210,
    "ts": 1514657508217.088, "dur": 20.361,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657508291.242, "dur": 108.868,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657508313.508, "dur": 83.204,
    "args": {
      "External id": 2487066,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657508323.252, "dur": 69.258,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657508412.816, "dur": 1.972,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514657508487.718, "dur": 1645.773,
    "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": 1336753, "tid": 1381210, "ts": 1514657508487.718,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1381210,
    "ts": 1514657508613.295, "dur": 100.377,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657508752.528, "dur": 41.670,
    "args": {
      "External id": 2487071,"kernel_hash": "c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/5t/c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr.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": 1336753, "tid": 1381210,
    "ts": 1514657508814.239, "dur": 46.952,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657508871.035, "dur": 33.643,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657508911.458, "dur": 45.392,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657508964.802, "dur": 66.952,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657509045.477, "dur": 46.200,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657509114.772, "dur": 24.183,
    "args": {
      "External id": 2487077,"kernel_hash": "cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/mm/cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k.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": 1336753, "tid": 1381210,
    "ts": 1514657509157.140, "dur": 28.883,
    "args": {
      "External id": 2487078,"kernel_hash": "c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2", "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/4b/c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2.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": 1336753, "tid": 1381210,
    "ts": 1514657509204.191, "dur": 21.529,
    "args": {
      "External id": 2487079,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657509239.483, "dur": 16.765,
    "args": {
      "External id": 2487080,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657509267.661, "dur": 29.598,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657509300.602, "dur": 31.991,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657509357.479, "dur": 192.033,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657509433.659, "dur": 5.867,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657509441.296, "dur": 2.458,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657509583.632, "dur": 26.916,
    "args": {
      "External id": 2487086,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657509622.026, "dur": 15.456,
    "args": {
      "External id": 2487087,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657509648.809, "dur": 38.914,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657509694.740, "dur": 35.551,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657509737.167, "dur": 23.188,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657509766.273, "dur": 30.815,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657509805.643, "dur": 22.026,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657509835.592, "dur": 30.060,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657509881.990, "dur": 23.650,
    "args": {
      "External id": 2487094,"kernel_hash": "cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/as/cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf.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": 1336753, "tid": 1381210,
    "ts": 1514657509936.537, "dur": 30.056,
    "args": {
      "External id": 2487095,"kernel_hash": "cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od", "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/ny/cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od.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": 1336753, "tid": 1381210,
    "ts": 1514657509979.991, "dur": 49.590,
    "args": {
      "External id": 2487096,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657510052.699, "dur": 17.310,
    "args": {
      "External id": 2487097,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657510082.986, "dur": 21.563,
    "args": {
      "External id": 2487098,"kernel_hash": "c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/25/c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6.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": 1336753, "tid": 1381210,
    "ts": 1514657510175.606, "dur": 15.660,
    "args": {
      "External id": 2487099,"Record function id": 0, "Ev Idx": 3898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657510179.670, "dur": 10.642,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510183.732, "dur": 5.622,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510185.353, "dur": 3.889,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510195.111, "dur": 5.612,
    "args": {
      "External id": 2487103,"Record function id": 0, "Ev Idx": 3902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657510196.604, "dur": 3.700,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510197.609, "dur": 2.207,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510198.701, "dur": 1.011,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510204.036, "dur": 4.799,
    "args": {
      "External id": 2487107,"Record function id": 0, "Ev Idx": 3906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657510205.349, "dur": 3.078,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510206.361, "dur": 1.615,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510206.890, "dur": 0.998,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510212.049, "dur": 4.519,
    "args": {
      "External id": 2487111,"Record function id": 0, "Ev Idx": 3910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657510213.421, "dur": 2.722,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510214.093, "dur": 1.638,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510214.875, "dur": 0.787,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510219.711, "dur": 3.543,
    "args": {
      "External id": 2487115,"Record function id": 0, "Ev Idx": 3914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657510220.770, "dur": 2.076,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510221.442, "dur": 1.018,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510221.717, "dur": 0.673,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510226.579, "dur": 4.325,
    "args": {
      "External id": 2487119,"Record function id": 0, "Ev Idx": 3918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657510227.815, "dur": 2.672,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510228.424, "dur": 1.563,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510229.041, "dur": 0.840,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510234.217, "dur": 5.461,
    "args": {
      "External id": 2487123,"Record function id": 0, "Ev Idx": 3922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657510235.268, "dur": 3.986,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510235.787, "dur": 3.000,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510236.070, "dur": 2.648,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510242.916, "dur": 3.624,
    "args": {
      "External id": 2487127,"Record function id": 0, "Ev Idx": 3926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657510243.988, "dur": 2.147,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510244.488, "dur": 1.212,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510244.783, "dur": 0.824,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510249.555, "dur": 29.717,
    "args": {
      "External id": 2487131,"Record function id": 0, "Ev Idx": 3930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657510250.681, "dur": 28.123,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510251.168, "dur": 27.061,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510277.307, "dur": 0.774,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510283.060, "dur": 36170.151,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510284.382, "dur": 36161.363,
    "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": 1336753, "tid": 1381210, "ts": 1514657510284.382,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.12)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657510313.133, "dur": 36.402,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510356.440, "dur": 61.931,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510423.994, "dur": 36014.790,
    "args": {
      "External id": 2487139,"Record function id": 0, "Ev Idx": 3938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657510504.525, "dur": 6.798,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510520.418, "dur": 4.795,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510558.158, "dur": 35116.402,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510571.651, "dur": 35093.987,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510632.556, "dur": 14.151,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510653.551, "dur": 34970.967,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510656.467, "dur": 34967.400,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510660.703, "dur": 5.803,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657510668.630, "dur": 34951.797,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657545759.985, "dur": 9.116,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657545763.301, "dur": 5.356,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657545800.303, "dur": 338.848,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657545825.942, "dur": 308.284,
    "args": {
      "External id": 2487152,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657545836.469, "dur": 291.850,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657546158.599, "dur": 1.976,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1381210,
    "ts": 1514657546218.046, "dur": 8.162,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657546275.659, "dur": 1.408,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657546294.825, "dur": 1.421,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657546309.009, "dur": 0.955,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657546321.292, "dur": 2.658,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657546335.716, "dur": 1.087,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657546349.440, "dur": 0.867,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657546363.646, "dur": 1.841,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657546376.868, "dur": 2.492,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657546467.667, "dur": 2684.061,
    "args": {
      "External id": 2487164,"Record function id": 0, "Ev Idx": 3963
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.11)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657546486.286, "dur": 982.602,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657546501.310, "dur": 322.792,
    "args": {
      "External id": 2487166,"Record function id": 0, "Ev Idx": 3965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1381210,
    "ts": 1514657546594.029, "dur": 4.674,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657546602.157, "dur": 1.174,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657546605.401, "dur": 1.058,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657546608.076, "dur": 0.904,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657546610.482, "dur": 0.770,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657546613.176, "dur": 0.735,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657546615.575, "dur": 3.453,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657546620.258, "dur": 1.292,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657546623.043, "dur": 1.135,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657546625.319, "dur": 0.673,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657546644.108, "dur": 146.116,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657546659.070, "dur": 126.972,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657546678.285, "dur": 12.056,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657546693.572, "dur": 63.701,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657546696.785, "dur": 60.132,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657546700.596, "dur": 5.561,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657546708.092, "dur": 48.204,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657546901.463, "dur": 558.785,
    "args": {
      "External id": 2487184,"Record function id": 0, "Ev Idx": 3983
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.10)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657546918.743, "dur": 528.483,
    "args": {
      "External id": 2487185,"Record function id": 0, "Ev Idx": 3984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657546972.169, "dur": 4.292,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657547027.869, "dur": 32.112,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657547033.005, "dur": 1.922,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657547037.522, "dur": 2.850,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657547041.895, "dur": 0.587,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657547044.583, "dur": 0.554,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657547046.811, "dur": 0.434,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657547048.669, "dur": 0.347,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657547050.610, "dur": 0.422,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657547053.411, "dur": 0.734,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657547055.568, "dur": 0.388,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657547070.527, "dur": 32.570,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657547136.621, "dur": 99.783,
    "args": {
      "External id": 2487198,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657547146.783, "dur": 5.990,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657547158.080, "dur": 9.809,
    "args": {
      "External id": 2487200,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657547162.162, "dur": 5.334,
    "args": {
      "External id": 2487201,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1381210,
    "ts": 1514657547165.908, "dur": 0.591,
    "args": {
      "External id": 2487202,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657547174.606, "dur": 24.844,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657547176.866, "dur": 0.369,
    "args": {
      "External id": 2487204,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657547178.887, "dur": 0.439,
    "args": {
      "External id": 2487205,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1381210,
    "ts": 1514657547181.094, "dur": 0.541,
    "args": {
      "External id": 2487206,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1381210,
    "ts": 1514657547183.502, "dur": 0.488,
    "args": {
      "External id": 2487207,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1381210,
    "ts": 1514657547185.271, "dur": 0.475,
    "args": {
      "External id": 2487208,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1381210,
    "ts": 1514657547187.479, "dur": 0.436,
    "args": {
      "External id": 2487209,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1381210,
    "ts": 1514657547189.355, "dur": 2.510,
    "args": {
      "External id": 2487210,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1381210,
    "ts": 1514657547193.594, "dur": 0.395,
    "args": {
      "External id": 2487211,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1381210,
    "ts": 1514657547195.722, "dur": 0.595,
    "args": {
      "External id": 2487212,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1381210,
    "ts": 1514657547208.861, "dur": 20.403,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657547280.876, "dur": 106.489,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657547301.908, "dur": 82.212,
    "args": {
      "External id": 2487215,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657547311.104, "dur": 69.020,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657547399.590, "dur": 1.694,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514657547475.250, "dur": 1656.214,
    "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": 1336753, "tid": 1381210, "ts": 1514657547475.250,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1381210,
    "ts": 1514657547604.089, "dur": 100.185,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657547743.262, "dur": 41.229,
    "args": {
      "External id": 2487220,"kernel_hash": "c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/5t/c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr.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": 1336753, "tid": 1381210,
    "ts": 1514657547803.688, "dur": 49.662,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657547863.281, "dur": 32.855,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657547903.216, "dur": 46.389,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657547959.380, "dur": 63.922,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657548035.297, "dur": 47.107,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657548105.939, "dur": 23.964,
    "args": {
      "External id": 2487226,"kernel_hash": "cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/mm/cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k.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": 1336753, "tid": 1381210,
    "ts": 1514657548147.179, "dur": 27.732,
    "args": {
      "External id": 2487227,"kernel_hash": "c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2", "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/4b/c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2.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": 1336753, "tid": 1381210,
    "ts": 1514657548193.469, "dur": 20.166,
    "args": {
      "External id": 2487228,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657548228.124, "dur": 16.816,
    "args": {
      "External id": 2487229,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657548256.003, "dur": 31.164,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657548290.783, "dur": 33.351,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657548349.649, "dur": 167.714,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657548424.309, "dur": 5.741,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657548431.944, "dur": 3.779,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657548568.098, "dur": 28.678,
    "args": {
      "External id": 2487235,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657548607.467, "dur": 15.904,
    "args": {
      "External id": 2487236,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657548632.404, "dur": 40.519,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657548679.935, "dur": 35.035,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657548721.719, "dur": 22.328,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657548751.880, "dur": 29.514,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657548788.095, "dur": 22.196,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657548817.109, "dur": 29.478,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657548862.592, "dur": 23.109,
    "args": {
      "External id": 2487243,"kernel_hash": "cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/as/cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf.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": 1336753, "tid": 1381210,
    "ts": 1514657548901.070, "dur": 23.724,
    "args": {
      "External id": 2487244,"kernel_hash": "cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od", "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/ny/cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od.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": 1336753, "tid": 1381210,
    "ts": 1514657548951.384, "dur": 23.900,
    "args": {
      "External id": 2487245,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657549042.835, "dur": 23.474,
    "args": {
      "External id": 2487246,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657549082.699, "dur": 17.050,
    "args": {
      "External id": 2487247,"kernel_hash": "c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/25/c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6.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": 1336753, "tid": 1381210,
    "ts": 1514657549173.423, "dur": 17.766,
    "args": {
      "External id": 2487248,"Record function id": 0, "Ev Idx": 4047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657549179.793, "dur": 10.456,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549183.497, "dur": 5.983,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549185.375, "dur": 3.987,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549195.240, "dur": 5.687,
    "args": {
      "External id": 2487252,"Record function id": 0, "Ev Idx": 4051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657549196.710, "dur": 3.779,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549197.801, "dur": 2.227,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549198.768, "dur": 1.167,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549204.182, "dur": 4.582,
    "args": {
      "External id": 2487256,"Record function id": 0, "Ev Idx": 4055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657549205.579, "dur": 2.776,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549206.354, "dur": 1.584,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549206.866, "dur": 1.002,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549212.124, "dur": 6.401,
    "args": {
      "External id": 2487260,"Record function id": 0, "Ev Idx": 4059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657549213.590, "dur": 4.525,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549214.148, "dur": 3.572,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549214.940, "dur": 2.708,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549221.721, "dur": 4.294,
    "args": {
      "External id": 2487264,"Record function id": 0, "Ev Idx": 4063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657549223.040, "dur": 2.561,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549223.621, "dur": 1.568,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549224.257, "dur": 0.870,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549229.295, "dur": 3.675,
    "args": {
      "External id": 2487268,"Record function id": 0, "Ev Idx": 4067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657549230.275, "dur": 2.235,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549230.728, "dur": 1.385,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549231.321, "dur": 0.685,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549239.379, "dur": 3.709,
    "args": {
      "External id": 2487272,"Record function id": 0, "Ev Idx": 4071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657549240.453, "dur": 2.207,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549241.212, "dur": 1.054,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549241.545, "dur": 0.615,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549246.243, "dur": 3.970,
    "args": {
      "External id": 2487276,"Record function id": 0, "Ev Idx": 4075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657549247.282, "dur": 2.511,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549248.124, "dur": 1.276,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549248.585, "dur": 0.740,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549253.427, "dur": 4.051,
    "args": {
      "External id": 2487280,"Record function id": 0, "Ev Idx": 4079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657549254.453, "dur": 2.600,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549255.007, "dur": 1.639,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549255.874, "dur": 0.673,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549261.060, "dur": 37764.518,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549262.409, "dur": 37718.699,
    "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": 1336753, "tid": 1381210, "ts": 1514657549262.409,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.11)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657549289.054, "dur": 35.598,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549332.223, "dur": 60.502,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549398.647, "dur": 37575.125,
    "args": {
      "External id": 2487288,"Record function id": 0, "Ev Idx": 4087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657549483.341, "dur": 6.435,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549498.799, "dur": 5.049,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549518.472, "dur": 36644.911,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549549.713, "dur": 36604.318,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549622.296, "dur": 15.508,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549644.242, "dur": 36469.810,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549646.744, "dur": 36466.524,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549650.841, "dur": 5.149,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657549727.963, "dur": 36382.055,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657586252.242, "dur": 8.358,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657586254.991, "dur": 5.216,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657586291.270, "dur": 382.143,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657586316.332, "dur": 352.172,
    "args": {
      "External id": 2487301,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657586326.713, "dur": 335.752,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657586696.477, "dur": 2.536,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1381210,
    "ts": 1514657586762.797, "dur": 6.289,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657586812.531, "dur": 1.337,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657586829.389, "dur": 1.361,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657586844.063, "dur": 0.926,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657586856.972, "dur": 1.090,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657586868.630, "dur": 1.221,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657586884.867, "dur": 1.037,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657586899.606, "dur": 1.933,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657586912.070, "dur": 0.873,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587042.146, "dur": 2770.490,
    "args": {
      "External id": 2487313,"Record function id": 0, "Ev Idx": 4112
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.10)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657587063.824, "dur": 1030.334,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587078.526, "dur": 325.329,
    "args": {
      "External id": 2487315,"Record function id": 0, "Ev Idx": 4114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1381210,
    "ts": 1514657587156.758, "dur": 4.872,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587164.745, "dur": 0.943,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587167.715, "dur": 0.715,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587170.246, "dur": 2.670,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587174.168, "dur": 1.172,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587177.248, "dur": 0.875,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587179.454, "dur": 1.695,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587182.541, "dur": 0.769,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587184.613, "dur": 1.113,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587187.120, "dur": 0.970,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587204.989, "dur": 164.672,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587220.457, "dur": 144.787,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587252.236, "dur": 12.302,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587267.568, "dur": 67.790,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587270.181, "dur": 64.793,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587274.589, "dur": 7.474,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587283.857, "dur": 50.468,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587481.898, "dur": 603.577,
    "args": {
      "External id": 2487333,"Record function id": 0, "Ev Idx": 4132
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.9)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657587498.021, "dur": 573.708,
    "args": {
      "External id": 2487334,"Record function id": 0, "Ev Idx": 4133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657587575.556, "dur": 6.267,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587598.593, "dur": 35.302,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587603.762, "dur": 1.709,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587607.965, "dur": 0.622,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587610.285, "dur": 0.525,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587612.547, "dur": 0.796,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587614.859, "dur": 0.453,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587616.921, "dur": 0.708,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587623.498, "dur": 3.035,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587628.018, "dur": 0.366,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587629.682, "dur": 0.329,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587643.994, "dur": 35.853,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587711.529, "dur": 97.941,
    "args": {
      "External id": 2487347,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657587721.666, "dur": 3.414,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587730.235, "dur": 10.056,
    "args": {
      "External id": 2487349,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657587734.578, "dur": 5.306,
    "args": {
      "External id": 2487350,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1381210,
    "ts": 1514657587738.372, "dur": 0.540,
    "args": {
      "External id": 2487351,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657587750.467, "dur": 22.903,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587752.716, "dur": 0.520,
    "args": {
      "External id": 2487353,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657587754.918, "dur": 0.373,
    "args": {
      "External id": 2487354,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1381210,
    "ts": 1514657587756.756, "dur": 0.442,
    "args": {
      "External id": 2487355,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1381210,
    "ts": 1514657587758.635, "dur": 2.441,
    "args": {
      "External id": 2487356,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1381210,
    "ts": 1514657587762.483, "dur": 0.364,
    "args": {
      "External id": 2487357,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1381210,
    "ts": 1514657587764.508, "dur": 0.407,
    "args": {
      "External id": 2487358,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1381210,
    "ts": 1514657587766.464, "dur": 0.596,
    "args": {
      "External id": 2487359,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1381210,
    "ts": 1514657587768.577, "dur": 0.336,
    "args": {
      "External id": 2487360,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1381210,
    "ts": 1514657587770.570, "dur": 0.364,
    "args": {
      "External id": 2487361,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1381210,
    "ts": 1514657587782.580, "dur": 19.696,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587854.390, "dur": 106.978,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587876.385, "dur": 81.481,
    "args": {
      "External id": 2487364,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657587885.675, "dur": 68.173,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657587975.946, "dur": 1.800,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514657588102.638, "dur": 1690.595,
    "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": 1336753, "tid": 1381210, "ts": 1514657588102.638,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1381210,
    "ts": 1514657588209.298, "dur": 100.507,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657588351.977, "dur": 41.764,
    "args": {
      "External id": 2487369,"kernel_hash": "c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/5t/c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr.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": 1336753, "tid": 1381210,
    "ts": 1514657588410.449, "dur": 48.457,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657588467.902, "dur": 32.666,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657588509.562, "dur": 65.060,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657588585.425, "dur": 31.863,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657588624.565, "dur": 43.114,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657588689.493, "dur": 24.874,
    "args": {
      "External id": 2487375,"kernel_hash": "cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/mm/cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k.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": 1336753, "tid": 1381210,
    "ts": 1514657588776.338, "dur": 29.867,
    "args": {
      "External id": 2487376,"kernel_hash": "c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2", "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/4b/c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2.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": 1336753, "tid": 1381210,
    "ts": 1514657588827.092, "dur": 23.957,
    "args": {
      "External id": 2487377,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657588865.351, "dur": 16.679,
    "args": {
      "External id": 2487378,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657588891.284, "dur": 35.913,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657588930.560, "dur": 36.121,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589032.695, "dur": 178.752,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589111.581, "dur": 6.643,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589120.411, "dur": 2.219,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589248.671, "dur": 24.739,
    "args": {
      "External id": 2487384,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657589283.943, "dur": 15.540,
    "args": {
      "External id": 2487385,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657589308.232, "dur": 40.675,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589354.984, "dur": 35.650,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589399.087, "dur": 32.677,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589442.722, "dur": 38.224,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589487.904, "dur": 22.636,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589518.072, "dur": 49.562,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589592.656, "dur": 25.841,
    "args": {
      "External id": 2487392,"kernel_hash": "cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/as/cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf.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": 1336753, "tid": 1381210,
    "ts": 1514657589638.122, "dur": 24.371,
    "args": {
      "External id": 2487393,"kernel_hash": "cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od", "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/ny/cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od.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": 1336753, "tid": 1381210,
    "ts": 1514657589676.876, "dur": 18.977,
    "args": {
      "External id": 2487394,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657589712.283, "dur": 19.613,
    "args": {
      "External id": 2487395,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657589744.810, "dur": 17.096,
    "args": {
      "External id": 2487396,"kernel_hash": "c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/25/c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6.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": 1336753, "tid": 1381210,
    "ts": 1514657589834.200, "dur": 18.320,
    "args": {
      "External id": 2487397,"Record function id": 0, "Ev Idx": 4196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657589837.548, "dur": 13.915,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589841.630, "dur": 5.772,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589843.298, "dur": 3.987,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589856.383, "dur": 5.052,
    "args": {
      "External id": 2487401,"Record function id": 0, "Ev Idx": 4200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657589857.479, "dur": 3.522,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589858.260, "dur": 2.261,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589859.411, "dur": 1.044,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589864.653, "dur": 6.389,
    "args": {
      "External id": 2487405,"Record function id": 0, "Ev Idx": 4204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657589865.781, "dur": 4.841,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589866.400, "dur": 3.765,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589867.218, "dur": 2.857,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589874.266, "dur": 3.942,
    "args": {
      "External id": 2487409,"Record function id": 0, "Ev Idx": 4208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657589875.408, "dur": 2.413,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589875.846, "dur": 1.568,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589876.634, "dur": 0.703,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589881.318, "dur": 4.125,
    "args": {
      "External id": 2487413,"Record function id": 0, "Ev Idx": 4212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657589882.676, "dur": 2.368,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589883.537, "dur": 1.108,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589883.971, "dur": 0.613,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589888.593, "dur": 4.033,
    "args": {
      "External id": 2487417,"Record function id": 0, "Ev Idx": 4216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657589889.675, "dur": 2.533,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589890.241, "dur": 1.507,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589890.914, "dur": 0.726,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589895.914, "dur": 3.422,
    "args": {
      "External id": 2487421,"Record function id": 0, "Ev Idx": 4220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657589896.893, "dur": 2.035,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589897.355, "dur": 1.180,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589897.778, "dur": 0.655,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589902.404, "dur": 3.515,
    "args": {
      "External id": 2487425,"Record function id": 0, "Ev Idx": 4224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657589903.342, "dur": 2.159,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589903.917, "dur": 1.171,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589904.214, "dur": 0.799,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589912.190, "dur": 3.992,
    "args": {
      "External id": 2487429,"Record function id": 0, "Ev Idx": 4228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657589913.220, "dur": 2.559,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589913.809, "dur": 1.564,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589914.489, "dur": 0.790,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589919.631, "dur": 34698.514,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657589921.069, "dur": 34688.973,
    "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": 1336753, "tid": 1381210, "ts": 1514657589921.069,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.10)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657589947.169, "dur": 71.464,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657590029.002, "dur": 63.394,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657590098.985, "dur": 34502.497,
    "args": {
      "External id": 2487437,"Record function id": 0, "Ev Idx": 4236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657590192.391, "dur": 7.452,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657590209.890, "dur": 7.161,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657590231.109, "dur": 33559.711,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657590244.318, "dur": 33537.895,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657590309.922, "dur": 13.690,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657590329.799, "dur": 33411.964,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657590332.428, "dur": 33408.436,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657590336.544, "dur": 5.363,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657590343.678, "dur": 33393.394,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657623878.326, "dur": 8.880,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657623881.346, "dur": 5.499,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657623917.478, "dur": 377.022,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657623943.630, "dur": 345.726,
    "args": {
      "External id": 2487450,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657623954.054, "dur": 329.267,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657624314.834, "dur": 2.289,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1381210,
    "ts": 1514657624375.586, "dur": 6.444,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657624426.053, "dur": 1.633,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657624443.472, "dur": 3.348,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657624459.804, "dur": 1.007,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657624472.788, "dur": 0.852,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657624484.505, "dur": 0.779,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657624497.316, "dur": 2.733,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657624511.331, "dur": 1.963,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657624523.062, "dur": 1.055,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657624633.610, "dur": 2707.697,
    "args": {
      "External id": 2487462,"Record function id": 0, "Ev Idx": 4261
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.9)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657624652.196, "dur": 1012.674,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657624665.161, "dur": 305.455,
    "args": {
      "External id": 2487464,"Record function id": 0, "Ev Idx": 4263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1381210,
    "ts": 1514657624739.499, "dur": 4.326,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657624746.826, "dur": 0.954,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657624749.723, "dur": 2.778,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657624754.196, "dur": 1.022,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657624756.636, "dur": 0.995,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657624759.220, "dur": 1.581,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657624762.472, "dur": 1.434,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657624765.235, "dur": 0.925,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657624767.547, "dur": 0.832,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657624769.763, "dur": 0.821,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657624787.030, "dur": 153.765,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657624801.937, "dur": 134.744,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657624824.100, "dur": 14.120,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657624841.493, "dur": 66.611,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657624844.216, "dur": 63.494,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657624848.489, "dur": 5.184,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657624857.134, "dur": 49.867,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657625093.312, "dur": 562.896,
    "args": {
      "External id": 2487482,"Record function id": 0, "Ev Idx": 4281
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.8)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657625110.943, "dur": 531.023,
    "args": {
      "External id": 2487483,"Record function id": 0, "Ev Idx": 4282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657625172.683, "dur": 5.869,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657625194.584, "dur": 28.605,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657625199.492, "dur": 1.702,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657625203.367, "dur": 0.427,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657625205.614, "dur": 0.737,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657625208.054, "dur": 0.407,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657625210.080, "dur": 0.643,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657625212.187, "dur": 2.458,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657625216.000, "dur": 0.422,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657625218.004, "dur": 0.387,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657625220.030, "dur": 0.370,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657625233.493, "dur": 34.176,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657625299.166, "dur": 100.939,
    "args": {
      "External id": 2487496,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657625309.327, "dur": 3.226,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657625317.918, "dur": 9.859,
    "args": {
      "External id": 2487498,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657625321.923, "dur": 5.443,
    "args": {
      "External id": 2487499,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1381210,
    "ts": 1514657625325.592, "dur": 0.734,
    "args": {
      "External id": 2487500,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657625334.922, "dur": 26.788,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657625337.288, "dur": 0.537,
    "args": {
      "External id": 2487502,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657625339.594, "dur": 0.370,
    "args": {
      "External id": 2487503,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1381210,
    "ts": 1514657625341.888, "dur": 2.203,
    "args": {
      "External id": 2487504,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1381210,
    "ts": 1514657625345.690, "dur": 0.593,
    "args": {
      "External id": 2487505,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1381210,
    "ts": 1514657625348.021, "dur": 0.602,
    "args": {
      "External id": 2487506,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1381210,
    "ts": 1514657625350.320, "dur": 0.478,
    "args": {
      "External id": 2487507,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1381210,
    "ts": 1514657625352.244, "dur": 0.444,
    "args": {
      "External id": 2487508,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1381210,
    "ts": 1514657625354.154, "dur": 0.712,
    "args": {
      "External id": 2487509,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1381210,
    "ts": 1514657625356.434, "dur": 0.586,
    "args": {
      "External id": 2487510,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1381210,
    "ts": 1514657625370.956, "dur": 21.432,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657625443.181, "dur": 129.510,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657625463.996, "dur": 104.784,
    "args": {
      "External id": 2487513,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657625472.895, "dur": 90.983,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657625590.136, "dur": 2.160,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514657625671.955, "dur": 1650.409,
    "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": 1336753, "tid": 1381210, "ts": 1514657625671.955,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1381210,
    "ts": 1514657625777.982, "dur": 101.151,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657625914.596, "dur": 40.586,
    "args": {
      "External id": 2487518,"kernel_hash": "c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/5t/c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr.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": 1336753, "tid": 1381210,
    "ts": 1514657625972.289, "dur": 93.750,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657626082.210, "dur": 36.104,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657626125.876, "dur": 45.821,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657626179.229, "dur": 29.511,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657626215.527, "dur": 42.614,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657626283.182, "dur": 24.085,
    "args": {
      "External id": 2487524,"kernel_hash": "cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/mm/cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k.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": 1336753, "tid": 1381210,
    "ts": 1514657626325.379, "dur": 29.325,
    "args": {
      "External id": 2487525,"kernel_hash": "c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2", "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/4b/c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2.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": 1336753, "tid": 1381210,
    "ts": 1514657626372.365, "dur": 19.904,
    "args": {
      "External id": 2487526,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657626405.249, "dur": 16.816,
    "args": {
      "External id": 2487527,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657626430.950, "dur": 30.533,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657626464.643, "dur": 31.914,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657626524.110, "dur": 187.483,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657626615.982, "dur": 6.932,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657626625.162, "dur": 2.699,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657626741.955, "dur": 25.467,
    "args": {
      "External id": 2487533,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657626778.124, "dur": 15.977,
    "args": {
      "External id": 2487534,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657626802.409, "dur": 39.101,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657626848.723, "dur": 34.414,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657626891.800, "dur": 23.250,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657626937.606, "dur": 32.542,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657626977.803, "dur": 58.573,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627047.410, "dur": 44.243,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627117.839, "dur": 30.819,
    "args": {
      "External id": 2487541,"kernel_hash": "cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/as/cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf.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": 1336753, "tid": 1381210,
    "ts": 1514657627165.272, "dur": 27.165,
    "args": {
      "External id": 2487542,"kernel_hash": "cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od", "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/ny/cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od.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": 1336753, "tid": 1381210,
    "ts": 1514657627207.775, "dur": 18.878,
    "args": {
      "External id": 2487543,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657627243.405, "dur": 19.017,
    "args": {
      "External id": 2487544,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657627274.599, "dur": 16.495,
    "args": {
      "External id": 2487545,"kernel_hash": "c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/25/c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6.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": 1336753, "tid": 1381210,
    "ts": 1514657627363.072, "dur": 14.633,
    "args": {
      "External id": 2487546,"Record function id": 0, "Ev Idx": 4345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657627366.180, "dur": 10.668,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627369.982, "dur": 6.026,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627371.924, "dur": 3.976,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627381.537, "dur": 5.596,
    "args": {
      "External id": 2487550,"Record function id": 0, "Ev Idx": 4349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657627383.206, "dur": 3.508,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627384.194, "dur": 2.065,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627384.986, "dur": 1.167,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627390.327, "dur": 4.816,
    "args": {
      "External id": 2487554,"Record function id": 0, "Ev Idx": 4353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657627391.664, "dur": 3.044,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627392.472, "dur": 1.812,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627393.149, "dur": 1.070,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627398.447, "dur": 6.753,
    "args": {
      "External id": 2487558,"Record function id": 0, "Ev Idx": 4357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657627399.720, "dur": 5.072,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627400.608, "dur": 3.788,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627401.228, "dur": 3.076,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627408.346, "dur": 4.140,
    "args": {
      "External id": 2487562,"Record function id": 0, "Ev Idx": 4361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657627409.750, "dur": 2.305,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627410.208, "dur": 1.431,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627410.628, "dur": 0.947,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627415.731, "dur": 3.953,
    "args": {
      "External id": 2487566,"Record function id": 0, "Ev Idx": 4365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657627417.131, "dur": 2.130,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627417.585, "dur": 1.283,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627418.184, "dur": 0.584,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627422.998, "dur": 3.667,
    "args": {
      "External id": 2487570,"Record function id": 0, "Ev Idx": 4369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657627424.252, "dur": 2.013,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627424.675, "dur": 1.202,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627424.975, "dur": 0.800,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627429.825, "dur": 4.038,
    "args": {
      "External id": 2487574,"Record function id": 0, "Ev Idx": 4373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657627431.102, "dur": 2.360,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627431.826, "dur": 1.257,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627432.385, "dur": 0.621,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627436.989, "dur": 4.077,
    "args": {
      "External id": 2487578,"Record function id": 0, "Ev Idx": 4377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657627438.413, "dur": 2.243,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627438.838, "dur": 1.414,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627439.347, "dur": 0.816,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627444.648, "dur": 35388.661,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627445.948, "dur": 35379.669,
    "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": 1336753, "tid": 1381210, "ts": 1514657627445.948,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.9)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657627472.428, "dur": 34.411,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627514.556, "dur": 84.054,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627606.131, "dur": 35211.465,
    "args": {
      "External id": 2487586,"Record function id": 0, "Ev Idx": 4385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657627692.864, "dur": 10.293,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627713.233, "dur": 4.784,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627732.306, "dur": 34186.293,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627745.506, "dur": 34163.971,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627800.627, "dur": 15.306,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627825.494, "dur": 34044.554,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627828.137, "dur": 34041.235,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627832.114, "dur": 4.958,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657627839.237, "dur": 34026.639,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657662047.413, "dur": 9.642,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657662050.658, "dur": 5.847,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657662146.822, "dur": 363.632,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657662172.955, "dur": 332.596,
    "args": {
      "External id": 2487599,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657662184.335, "dur": 314.550,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657662543.539, "dur": 3.222,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1381210,
    "ts": 1514657662609.993, "dur": 6.761,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657662660.204, "dur": 1.441,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657662678.344, "dur": 1.034,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657662693.713, "dur": 1.232,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657662706.837, "dur": 1.078,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657662718.111, "dur": 1.197,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657662731.697, "dur": 1.010,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657662746.013, "dur": 1.309,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657662756.805, "dur": 0.814,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657662848.662, "dur": 2676.953,
    "args": {
      "External id": 2487611,"Record function id": 0, "Ev Idx": 4410
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.8)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657662867.449, "dur": 993.223,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657662880.932, "dur": 342.404,
    "args": {
      "External id": 2487613,"Record function id": 0, "Ev Idx": 4412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1381210,
    "ts": 1514657662954.674, "dur": 4.249,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657662962.040, "dur": 1.009,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657662964.613, "dur": 1.219,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657662967.954, "dur": 3.517,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657662972.953, "dur": 1.210,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657662975.834, "dur": 1.059,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657662978.588, "dur": 0.986,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657663015.230, "dur": 1.635,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657663020.483, "dur": 1.195,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657663023.192, "dur": 1.258,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657663043.579, "dur": 150.771,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657663059.578, "dur": 130.657,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657663079.472, "dur": 12.135,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657663096.745, "dur": 65.693,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657663099.374, "dur": 62.687,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657663103.059, "dur": 7.163,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657663112.111, "dur": 49.420,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657663303.615, "dur": 549.471,
    "args": {
      "External id": 2487631,"Record function id": 0, "Ev Idx": 4430
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.7)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657663320.217, "dur": 521.281,
    "args": {
      "External id": 2487632,"Record function id": 0, "Ev Idx": 4431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657663373.427, "dur": 4.271,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657663393.260, "dur": 35.298,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657663401.602, "dur": 1.559,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657663405.258, "dur": 0.596,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657663407.868, "dur": 0.842,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657663410.154, "dur": 0.426,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657663412.009, "dur": 0.840,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657663414.573, "dur": 0.740,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657663417.111, "dur": 2.519,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657663421.050, "dur": 0.408,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657663422.960, "dur": 0.411,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657663438.181, "dur": 31.376,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657663498.949, "dur": 122.123,
    "args": {
      "External id": 2487645,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657663508.842, "dur": 2.902,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657663516.785, "dur": 10.379,
    "args": {
      "External id": 2487647,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657663521.130, "dur": 5.591,
    "args": {
      "External id": 2487648,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1381210,
    "ts": 1514657663524.619, "dur": 0.676,
    "args": {
      "External id": 2487649,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657663551.850, "dur": 25.449,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657663554.848, "dur": 0.692,
    "args": {
      "External id": 2487651,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657663557.149, "dur": 0.915,
    "args": {
      "External id": 2487652,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1381210,
    "ts": 1514657663559.860, "dur": 0.581,
    "args": {
      "External id": 2487653,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1381210,
    "ts": 1514657663562.079, "dur": 2.176,
    "args": {
      "External id": 2487654,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1381210,
    "ts": 1514657663565.599, "dur": 0.829,
    "args": {
      "External id": 2487655,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1381210,
    "ts": 1514657663568.058, "dur": 0.821,
    "args": {
      "External id": 2487656,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1381210,
    "ts": 1514657663570.627, "dur": 0.253,
    "args": {
      "External id": 2487657,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1381210,
    "ts": 1514657663572.266, "dur": 0.400,
    "args": {
      "External id": 2487658,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1381210,
    "ts": 1514657663573.852, "dur": 0.420,
    "args": {
      "External id": 2487659,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1381210,
    "ts": 1514657663591.467, "dur": 21.915,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657663663.935, "dur": 113.329,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657663692.435, "dur": 81.552,
    "args": {
      "External id": 2487662,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657663700.881, "dur": 69.166,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657663793.715, "dur": 1.721,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514657663867.658, "dur": 1638.486,
    "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": 1336753, "tid": 1381210, "ts": 1514657663867.658,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1381210,
    "ts": 1514657663971.221, "dur": 140.593,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657664159.481, "dur": 39.198,
    "args": {
      "External id": 2487667,"kernel_hash": "c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/5t/c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr.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": 1336753, "tid": 1381210,
    "ts": 1514657664215.455, "dur": 48.657,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657664273.199, "dur": 31.930,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657664313.399, "dur": 43.713,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657664363.403, "dur": 27.388,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657664397.607, "dur": 41.187,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657664461.828, "dur": 22.842,
    "args": {
      "External id": 2487673,"kernel_hash": "cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/mm/cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k.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": 1336753, "tid": 1381210,
    "ts": 1514657664504.573, "dur": 46.044,
    "args": {
      "External id": 2487674,"kernel_hash": "c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2", "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/4b/c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2.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": 1336753, "tid": 1381210,
    "ts": 1514657664577.450, "dur": 22.125,
    "args": {
      "External id": 2487675,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657664614.886, "dur": 16.267,
    "args": {
      "External id": 2487676,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657664640.807, "dur": 33.190,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657664677.289, "dur": 32.296,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657664737.607, "dur": 168.688,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657664814.179, "dur": 6.292,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657664822.739, "dur": 1.988,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657664935.305, "dur": 25.194,
    "args": {
      "External id": 2487682,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657664971.997, "dur": 54.195,
    "args": {
      "External id": 2487683,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657665038.634, "dur": 50.182,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665096.185, "dur": 36.240,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665139.805, "dur": 22.960,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665170.058, "dur": 30.401,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665206.755, "dur": 21.351,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665238.342, "dur": 29.884,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665286.648, "dur": 28.096,
    "args": {
      "External id": 2487690,"kernel_hash": "cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/as/cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf.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": 1336753, "tid": 1381210,
    "ts": 1514657665339.826, "dur": 32.550,
    "args": {
      "External id": 2487691,"kernel_hash": "cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od", "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/ny/cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od.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": 1336753, "tid": 1381210,
    "ts": 1514657665390.690, "dur": 18.526,
    "args": {
      "External id": 2487692,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657665423.001, "dur": 19.158,
    "args": {
      "External id": 2487693,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657665456.533, "dur": 16.451,
    "args": {
      "External id": 2487694,"kernel_hash": "c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/25/c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6.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": 1336753, "tid": 1381210,
    "ts": 1514657665568.175, "dur": 15.643,
    "args": {
      "External id": 2487695,"Record function id": 0, "Ev Idx": 4494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657665571.560, "dur": 11.202,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665575.667, "dur": 5.886,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665577.195, "dur": 4.140,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665587.952, "dur": 5.508,
    "args": {
      "External id": 2487699,"Record function id": 0, "Ev Idx": 4498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657665589.418, "dur": 3.587,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665590.434, "dur": 2.083,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665591.408, "dur": 1.023,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665596.708, "dur": 5.867,
    "args": {
      "External id": 2487703,"Record function id": 0, "Ev Idx": 4502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657665598.182, "dur": 3.994,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665599.740, "dur": 1.974,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665600.332, "dur": 1.312,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665605.810, "dur": 4.019,
    "args": {
      "External id": 2487707,"Record function id": 0, "Ev Idx": 4506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657665607.144, "dur": 2.299,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665607.653, "dur": 1.400,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665608.243, "dur": 0.745,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665616.545, "dur": 6.400,
    "args": {
      "External id": 2487711,"Record function id": 0, "Ev Idx": 4510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657665617.912, "dur": 4.645,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665618.774, "dur": 3.265,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665619.112, "dur": 2.863,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665626.164, "dur": 4.063,
    "args": {
      "External id": 2487715,"Record function id": 0, "Ev Idx": 4514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657665627.347, "dur": 2.460,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665627.846, "dur": 1.520,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665628.357, "dur": 0.914,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665633.582, "dur": 4.146,
    "args": {
      "External id": 2487719,"Record function id": 0, "Ev Idx": 4518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657665634.628, "dur": 2.707,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665635.163, "dur": 1.732,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665635.978, "dur": 0.820,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665640.935, "dur": 4.119,
    "args": {
      "External id": 2487723,"Record function id": 0, "Ev Idx": 4522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657665642.154, "dur": 2.463,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665642.821, "dur": 1.323,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665643.299, "dur": 0.741,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665648.211, "dur": 4.621,
    "args": {
      "External id": 2487727,"Record function id": 0, "Ev Idx": 4526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657665649.688, "dur": 2.755,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665650.530, "dur": 1.492,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665651.268, "dur": 0.636,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665656.328, "dur": 35772.792,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665657.580, "dur": 35763.617,
    "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": 1336753, "tid": 1381210, "ts": 1514657665657.580,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.8)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657665690.268, "dur": 37.025,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665739.296, "dur": 70.075,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665815.441, "dur": 35598.446,
    "args": {
      "External id": 2487735,"Record function id": 0, "Ev Idx": 4534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657665904.518, "dur": 6.980,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665923.652, "dur": 4.709,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665942.282, "dur": 34638.063,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657665955.032, "dur": 34616.009,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657666053.108, "dur": 14.754,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657666074.182, "dur": 34448.230,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657666076.923, "dur": 34444.875,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657666081.314, "dur": 7.969,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657666091.099, "dur": 34427.730,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657700666.418, "dur": 10.334,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657700671.187, "dur": 5.225,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657700705.821, "dur": 416.048,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657700733.583, "dur": 382.791,
    "args": {
      "External id": 2487748,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657700743.795, "dur": 366.457,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657701146.052, "dur": 2.144,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1381210,
    "ts": 1514657701208.819, "dur": 6.886,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657701256.421, "dur": 1.319,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657701273.432, "dur": 1.419,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657701287.208, "dur": 3.223,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657701306.924, "dur": 1.005,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657701318.503, "dur": 0.878,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657701330.356, "dur": 0.825,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657701341.870, "dur": 2.462,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657701354.342, "dur": 0.866,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657701442.855, "dur": 2772.816,
    "args": {
      "External id": 2487760,"Record function id": 0, "Ev Idx": 4559
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.7)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657701461.625, "dur": 1097.173,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657701475.493, "dur": 324.000,
    "args": {
      "External id": 2487762,"Record function id": 0, "Ev Idx": 4561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1381210,
    "ts": 1514657701570.055, "dur": 4.882,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657701579.869, "dur": 0.883,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657701582.430, "dur": 0.959,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657701585.258, "dur": 0.955,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657701587.982, "dur": 0.739,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657701591.679, "dur": 3.273,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657701596.465, "dur": 0.659,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657701598.514, "dur": 1.136,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657701600.992, "dur": 0.998,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657701604.413, "dur": 1.162,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657701624.658, "dur": 145.568,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657701641.304, "dur": 124.551,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657701657.426, "dur": 13.471,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657701673.850, "dur": 64.905,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657701676.685, "dur": 61.649,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657701680.592, "dur": 6.802,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657701689.006, "dur": 48.767,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657701876.095, "dur": 671.651,
    "args": {
      "External id": 2487780,"Record function id": 0, "Ev Idx": 4579
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.6)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657701893.825, "dur": 622.393,
    "args": {
      "External id": 2487781,"Record function id": 0, "Ev Idx": 4580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657701948.557, "dur": 4.394,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657701972.312, "dur": 69.861,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657701976.726, "dur": 4.121,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657702018.569, "dur": 0.592,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657702020.961, "dur": 0.359,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657702024.078, "dur": 0.468,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657702026.038, "dur": 0.444,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657702027.973, "dur": 0.715,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657702031.073, "dur": 0.502,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657702033.221, "dur": 0.570,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657702035.362, "dur": 2.710,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657702053.906, "dur": 35.440,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657702122.837, "dur": 167.596,
    "args": {
      "External id": 2487794,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657702132.981, "dur": 4.947,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657702143.083, "dur": 11.548,
    "args": {
      "External id": 2487796,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657702148.816, "dur": 5.372,
    "args": {
      "External id": 2487797,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1381210,
    "ts": 1514657702152.679, "dur": 0.478,
    "args": {
      "External id": 2487798,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657702162.055, "dur": 24.919,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657702164.657, "dur": 0.764,
    "args": {
      "External id": 2487800,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657702167.468, "dur": 0.310,
    "args": {
      "External id": 2487801,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1381210,
    "ts": 1514657702169.249, "dur": 0.639,
    "args": {
      "External id": 2487802,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1381210,
    "ts": 1514657702171.112, "dur": 0.566,
    "args": {
      "External id": 2487803,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1381210,
    "ts": 1514657702173.497, "dur": 0.595,
    "args": {
      "External id": 2487804,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1381210,
    "ts": 1514657702175.459, "dur": 2.370,
    "args": {
      "External id": 2487805,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1381210,
    "ts": 1514657702179.263, "dur": 0.682,
    "args": {
      "External id": 2487806,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1381210,
    "ts": 1514657702181.292, "dur": 0.582,
    "args": {
      "External id": 2487807,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1381210,
    "ts": 1514657702183.110, "dur": 0.545,
    "args": {
      "External id": 2487808,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1381210,
    "ts": 1514657702260.663, "dur": 21.855,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657702338.437, "dur": 113.898,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657702365.491, "dur": 83.498,
    "args": {
      "External id": 2487811,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657702375.122, "dur": 69.579,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657702466.727, "dur": 1.821,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514657702566.751, "dur": 1629.204,
    "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": 1336753, "tid": 1381210, "ts": 1514657702566.751,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1381210,
    "ts": 1514657702673.734, "dur": 101.884,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657702818.365, "dur": 39.988,
    "args": {
      "External id": 2487816,"kernel_hash": "c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/5t/c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr.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": 1336753, "tid": 1381210,
    "ts": 1514657702873.644, "dur": 46.827,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657702929.821, "dur": 32.886,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657702973.091, "dur": 86.009,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657703071.228, "dur": 28.807,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657703107.085, "dur": 40.087,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657703172.254, "dur": 23.075,
    "args": {
      "External id": 2487822,"kernel_hash": "cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/mm/cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k.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": 1336753, "tid": 1381210,
    "ts": 1514657703215.088, "dur": 28.766,
    "args": {
      "External id": 2487823,"kernel_hash": "c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2", "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/4b/c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2.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": 1336753, "tid": 1381210,
    "ts": 1514657703263.352, "dur": 21.813,
    "args": {
      "External id": 2487824,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657703301.887, "dur": 16.420,
    "args": {
      "External id": 2487825,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657703326.491, "dur": 29.701,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657703359.267, "dur": 30.953,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657703415.240, "dur": 185.283,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657703490.257, "dur": 8.191,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657703500.002, "dur": 2.383,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657703635.700, "dur": 26.467,
    "args": {
      "External id": 2487831,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657703673.652, "dur": 15.358,
    "args": {
      "External id": 2487832,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657703697.033, "dur": 43.222,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657703745.919, "dur": 34.446,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657703786.428, "dur": 21.391,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657703813.429, "dur": 28.621,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657703846.764, "dur": 21.116,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657703874.876, "dur": 28.792,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657703923.029, "dur": 36.935,
    "args": {
      "External id": 2487839,"kernel_hash": "cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/as/cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf.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": 1336753, "tid": 1381210,
    "ts": 1514657704030.317, "dur": 28.229,
    "args": {
      "External id": 2487840,"kernel_hash": "cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od", "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/ny/cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od.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": 1336753, "tid": 1381210,
    "ts": 1514657704079.392, "dur": 19.179,
    "args": {
      "External id": 2487841,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657704117.377, "dur": 19.273,
    "args": {
      "External id": 2487842,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657704151.193, "dur": 16.787,
    "args": {
      "External id": 2487843,"kernel_hash": "c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/25/c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6.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": 1336753, "tid": 1381210,
    "ts": 1514657704237.658, "dur": 15.152,
    "args": {
      "External id": 2487844,"Record function id": 0, "Ev Idx": 4643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657704241.130, "dur": 10.774,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704245.512, "dur": 5.401,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704246.732, "dur": 4.057,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704256.659, "dur": 4.458,
    "args": {
      "External id": 2487848,"Record function id": 0, "Ev Idx": 4647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657704258.217, "dur": 2.482,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704258.828, "dur": 1.427,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704259.415, "dur": 0.771,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704264.329, "dur": 4.034,
    "args": {
      "External id": 2487852,"Record function id": 0, "Ev Idx": 4651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657704265.619, "dur": 2.307,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704266.263, "dur": 1.241,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704266.609, "dur": 0.830,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704271.519, "dur": 3.908,
    "args": {
      "External id": 2487856,"Record function id": 0, "Ev Idx": 4655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657704272.628, "dur": 2.328,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704273.199, "dur": 1.330,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704273.748, "dur": 0.668,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704278.510, "dur": 6.346,
    "args": {
      "External id": 2487860,"Record function id": 0, "Ev Idx": 4659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657704279.780, "dur": 4.660,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704280.341, "dur": 3.681,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704280.816, "dur": 3.105,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704287.989, "dur": 3.883,
    "args": {
      "External id": 2487864,"Record function id": 0, "Ev Idx": 4663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657704289.381, "dur": 2.074,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704289.876, "dur": 1.152,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704290.152, "dur": 0.789,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704295.152, "dur": 3.798,
    "args": {
      "External id": 2487868,"Record function id": 0, "Ev Idx": 4667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657704296.409, "dur": 2.111,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704296.901, "dur": 1.187,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704297.309, "dur": 0.694,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704302.213, "dur": 4.219,
    "args": {
      "External id": 2487872,"Record function id": 0, "Ev Idx": 4671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657704303.628, "dur": 2.315,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704304.213, "dur": 1.246,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704304.641, "dur": 0.715,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704310.094, "dur": 6.193,
    "args": {
      "External id": 2487876,"Record function id": 0, "Ev Idx": 4675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657704311.441, "dur": 4.308,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704312.148, "dur": 3.141,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704314.366, "dur": 0.835,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704320.459, "dur": 35609.822,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704321.794, "dur": 35600.937,
    "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": 1336753, "tid": 1381210, "ts": 1514657704321.794,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.7)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657704360.954, "dur": 39.300,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704408.647, "dur": 62.331,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704476.999, "dur": 35438.642,
    "args": {
      "External id": 2487884,"Record function id": 0, "Ev Idx": 4683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657704581.703, "dur": 7.567,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704599.177, "dur": 8.316,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704622.748, "dur": 34529.414,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704636.338, "dur": 34507.573,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704691.091, "dur": 13.107,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704710.343, "dur": 34396.447,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704713.092, "dur": 34393.026,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704717.090, "dur": 9.571,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657704728.320, "dur": 34374.555,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657739236.414, "dur": 9.106,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657739238.974, "dur": 6.256,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657739274.335, "dur": 356.869,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657739301.267, "dur": 325.077,
    "args": {
      "External id": 2487897,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657739311.345, "dur": 309.212,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657739653.599, "dur": 2.147,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1381210,
    "ts": 1514657739714.749, "dur": 6.366,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657739763.860, "dur": 1.521,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657739781.402, "dur": 1.841,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657739795.333, "dur": 2.840,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657739811.573, "dur": 1.310,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657739822.992, "dur": 0.634,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657739834.041, "dur": 0.809,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657739845.936, "dur": 2.342,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657739857.938, "dur": 0.763,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657739943.380, "dur": 2711.340,
    "args": {
      "External id": 2487909,"Record function id": 0, "Ev Idx": 4708
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.6)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657739960.961, "dur": 1083.627,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657739974.357, "dur": 355.664,
    "args": {
      "External id": 2487911,"Record function id": 0, "Ev Idx": 4710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1381210,
    "ts": 1514657740084.871, "dur": 4.753,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740093.044, "dur": 0.871,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740097.036, "dur": 0.963,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740099.939, "dur": 0.769,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740102.170, "dur": 0.817,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740104.367, "dur": 2.911,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740109.798, "dur": 0.785,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740112.056, "dur": 0.902,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740114.268, "dur": 0.723,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740116.746, "dur": 0.643,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740137.029, "dur": 161.920,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740153.085, "dur": 141.541,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740184.008, "dur": 12.224,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740199.267, "dur": 67.389,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740201.730, "dur": 64.523,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740208.271, "dur": 6.114,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740215.869, "dur": 49.628,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740409.556, "dur": 626.225,
    "args": {
      "External id": 2487929,"Record function id": 0, "Ev Idx": 4728
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.5)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657740425.446, "dur": 593.200,
    "args": {
      "External id": 2487930,"Record function id": 0, "Ev Idx": 4729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657740480.000, "dur": 4.609,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740504.074, "dur": 47.784,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740508.110, "dur": 3.940,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740514.120, "dur": 0.449,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740517.036, "dur": 0.257,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740518.931, "dur": 0.378,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740520.771, "dur": 1.175,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740523.311, "dur": 0.521,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740525.230, "dur": 0.405,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740542.935, "dur": 0.710,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740546.330, "dur": 2.501,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740562.869, "dur": 34.085,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740629.228, "dur": 132.805,
    "args": {
      "External id": 2487943,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657740639.485, "dur": 3.905,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740648.282, "dur": 9.436,
    "args": {
      "External id": 2487945,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657740652.179, "dur": 5.148,
    "args": {
      "External id": 2487946,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1381210,
    "ts": 1514657740655.839, "dur": 0.409,
    "args": {
      "External id": 2487947,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657740664.478, "dur": 28.017,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740666.687, "dur": 0.404,
    "args": {
      "External id": 2487949,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657740669.042, "dur": 0.428,
    "args": {
      "External id": 2487950,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1381210,
    "ts": 1514657740673.685, "dur": 0.310,
    "args": {
      "External id": 2487951,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1381210,
    "ts": 1514657740675.332, "dur": 0.504,
    "args": {
      "External id": 2487952,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1381210,
    "ts": 1514657740677.184, "dur": 0.266,
    "args": {
      "External id": 2487953,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1381210,
    "ts": 1514657740678.764, "dur": 2.110,
    "args": {
      "External id": 2487954,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1381210,
    "ts": 1514657740682.323, "dur": 0.624,
    "args": {
      "External id": 2487955,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1381210,
    "ts": 1514657740684.550, "dur": 0.476,
    "args": {
      "External id": 2487956,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1381210,
    "ts": 1514657740686.263, "dur": 0.459,
    "args": {
      "External id": 2487957,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1381210,
    "ts": 1514657740733.195, "dur": 21.324,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740808.857, "dur": 111.311,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740833.953, "dur": 82.848,
    "args": {
      "External id": 2487960,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657740843.282, "dur": 69.540,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657740934.048, "dur": 1.889,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514657741052.052, "dur": 1582.875,
    "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": 1336753, "tid": 1381210, "ts": 1514657741052.052,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1381210,
    "ts": 1514657741161.144, "dur": 101.108,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657741304.351, "dur": 38.973,
    "args": {
      "External id": 2487965,"kernel_hash": "c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/5t/c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr.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": 1336753, "tid": 1381210,
    "ts": 1514657741359.354, "dur": 47.419,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657741415.609, "dur": 31.705,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657741453.647, "dur": 43.007,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657741504.908, "dur": 46.218,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657741562.080, "dur": 44.679,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657741630.830, "dur": 22.997,
    "args": {
      "External id": 2487971,"kernel_hash": "cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/mm/cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k.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": 1336753, "tid": 1381210,
    "ts": 1514657741672.199, "dur": 27.249,
    "args": {
      "External id": 2487972,"kernel_hash": "c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2", "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/4b/c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2.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": 1336753, "tid": 1381210,
    "ts": 1514657741717.839, "dur": 20.243,
    "args": {
      "External id": 2487973,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657741754.245, "dur": 15.635,
    "args": {
      "External id": 2487974,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657741777.416, "dur": 29.515,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657741809.989, "dur": 31.351,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657741870.152, "dur": 201.723,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657741944.492, "dur": 7.702,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657741953.900, "dur": 2.188,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742107.449, "dur": 25.194,
    "args": {
      "External id": 2487980,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657742144.420, "dur": 14.944,
    "args": {
      "External id": 2487981,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657742167.203, "dur": 39.492,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742211.927, "dur": 33.680,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742251.518, "dur": 21.652,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742279.073, "dur": 28.598,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742312.889, "dur": 20.697,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742340.422, "dur": 28.119,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742386.700, "dur": 28.978,
    "args": {
      "External id": 2487988,"kernel_hash": "cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/as/cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf.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": 1336753, "tid": 1381210,
    "ts": 1514657742432.707, "dur": 27.122,
    "args": {
      "External id": 2487989,"kernel_hash": "cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od", "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/ny/cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od.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": 1336753, "tid": 1381210,
    "ts": 1514657742472.814, "dur": 17.102,
    "args": {
      "External id": 2487990,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657742508.429, "dur": 42.744,
    "args": {
      "External id": 2487991,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657742581.124, "dur": 21.600,
    "args": {
      "External id": 2487992,"kernel_hash": "c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/25/c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6.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": 1336753, "tid": 1381210,
    "ts": 1514657742676.326, "dur": 17.873,
    "args": {
      "External id": 2487993,"Record function id": 0, "Ev Idx": 4792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657742679.761, "dur": 13.541,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742687.131, "dur": 5.227,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742688.525, "dur": 3.751,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742698.153, "dur": 4.639,
    "args": {
      "External id": 2487997,"Record function id": 0, "Ev Idx": 4796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657742699.466, "dur": 2.868,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742700.478, "dur": 1.449,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742701.142, "dur": 0.703,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742706.542, "dur": 4.659,
    "args": {
      "External id": 2488001,"Record function id": 0, "Ev Idx": 4800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657742707.962, "dur": 2.807,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742708.979, "dur": 1.348,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742709.626, "dur": 0.619,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742714.510, "dur": 3.656,
    "args": {
      "External id": 2488005,"Record function id": 0, "Ev Idx": 4804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657742715.852, "dur": 1.910,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742716.324, "dur": 1.013,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742716.646, "dur": 0.622,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742721.361, "dur": 5.569,
    "args": {
      "External id": 2488009,"Record function id": 0, "Ev Idx": 4808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657742722.437, "dur": 4.052,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742722.867, "dur": 3.226,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742723.156, "dur": 2.870,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742730.082, "dur": 4.709,
    "args": {
      "External id": 2488013,"Record function id": 0, "Ev Idx": 4812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657742731.372, "dur": 2.990,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742731.810, "dur": 2.143,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742732.995, "dur": 0.843,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742738.162, "dur": 6.551,
    "args": {
      "External id": 2488017,"Record function id": 0, "Ev Idx": 4816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657742739.239, "dur": 5.047,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742739.875, "dur": 3.981,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742740.273, "dur": 0.670,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742747.866, "dur": 3.368,
    "args": {
      "External id": 2488021,"Record function id": 0, "Ev Idx": 4820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657742748.968, "dur": 1.864,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742749.390, "dur": 1.003,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742749.683, "dur": 0.647,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742754.387, "dur": 4.063,
    "args": {
      "External id": 2488025,"Record function id": 0, "Ev Idx": 4824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657742755.462, "dur": 2.583,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742756.257, "dur": 1.400,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742756.917, "dur": 0.676,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742762.060, "dur": 36159.005,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742763.467, "dur": 36149.798,
    "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": 1336753, "tid": 1381210, "ts": 1514657742763.467,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.6)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657742790.046, "dur": 36.777,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742833.673, "dur": 62.962,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657742902.530, "dur": 36003.500,
    "args": {
      "External id": 2488033,"Record function id": 0, "Ev Idx": 4832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657743021.061, "dur": 8.674,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657743044.584, "dur": 4.901,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657743064.650, "dur": 35079.914,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657743077.713, "dur": 35058.047,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657743145.155, "dur": 16.739,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657743167.900, "dur": 34929.368,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657743170.378, "dur": 34926.077,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657743174.418, "dur": 5.592,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657743181.927, "dur": 34911.318,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657778225.654, "dur": 7.930,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657778228.383, "dur": 4.886,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657778263.787, "dur": 354.595,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657778290.732, "dur": 322.773,
    "args": {
      "External id": 2488046,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657778300.571, "dur": 307.021,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657778641.649, "dur": 2.289,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1381210,
    "ts": 1514657778702.129, "dur": 8.094,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657778753.546, "dur": 1.598,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657778770.502, "dur": 1.161,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657778784.244, "dur": 0.830,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657778796.648, "dur": 2.686,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657778810.278, "dur": 1.037,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657778821.804, "dur": 1.022,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657778833.348, "dur": 1.159,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657778845.746, "dur": 2.612,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657778935.239, "dur": 2639.551,
    "args": {
      "External id": 2488058,"Record function id": 0, "Ev Idx": 4857
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.5)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657778953.147, "dur": 984.438,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657778967.741, "dur": 335.764,
    "args": {
      "External id": 2488060,"Record function id": 0, "Ev Idx": 4859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1381210,
    "ts": 1514657779078.965, "dur": 4.894,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779086.845, "dur": 0.797,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779089.613, "dur": 0.789,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779092.332, "dur": 0.854,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779095.706, "dur": 0.772,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779098.088, "dur": 0.677,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779100.565, "dur": 0.667,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779102.766, "dur": 2.561,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779107.531, "dur": 1.012,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779110.008, "dur": 1.394,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779129.574, "dur": 145.450,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779145.683, "dur": 125.007,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779163.109, "dur": 12.092,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779178.022, "dur": 64.926,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779180.486, "dur": 62.084,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779184.328, "dur": 7.500,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779194.628, "dur": 47.317,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779377.383, "dur": 552.488,
    "args": {
      "External id": 2488078,"Record function id": 0, "Ev Idx": 4877
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.4)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657779393.438, "dur": 524.017,
    "args": {
      "External id": 2488079,"Record function id": 0, "Ev Idx": 4878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657779447.755, "dur": 4.612,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779470.851, "dur": 28.440,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779475.721, "dur": 1.552,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779479.984, "dur": 0.498,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779482.103, "dur": 2.765,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779486.265, "dur": 0.482,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779488.137, "dur": 0.388,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779490.578, "dur": 0.496,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779492.581, "dur": 0.352,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779494.509, "dur": 0.527,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779496.553, "dur": 0.373,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779508.403, "dur": 46.910,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779587.742, "dur": 104.437,
    "args": {
      "External id": 2488092,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657779597.831, "dur": 4.178,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779607.141, "dur": 12.117,
    "args": {
      "External id": 2488094,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657779610.999, "dur": 7.828,
    "args": {
      "External id": 2488095,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1381210,
    "ts": 1514657779614.989, "dur": 2.357,
    "args": {
      "External id": 2488096,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657779626.789, "dur": 24.993,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779628.934, "dur": 0.567,
    "args": {
      "External id": 2488098,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657779631.130, "dur": 0.304,
    "args": {
      "External id": 2488099,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1381210,
    "ts": 1514657779632.889, "dur": 0.442,
    "args": {
      "External id": 2488100,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1381210,
    "ts": 1514657779634.929, "dur": 0.462,
    "args": {
      "External id": 2488101,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1381210,
    "ts": 1514657779636.613, "dur": 0.532,
    "args": {
      "External id": 2488102,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1381210,
    "ts": 1514657779638.713, "dur": 0.563,
    "args": {
      "External id": 2488103,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1381210,
    "ts": 1514657779640.554, "dur": 0.418,
    "args": {
      "External id": 2488104,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1381210,
    "ts": 1514657779642.680, "dur": 2.006,
    "args": {
      "External id": 2488105,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1381210,
    "ts": 1514657779645.865, "dur": 0.335,
    "args": {
      "External id": 2488106,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1381210,
    "ts": 1514657779662.354, "dur": 22.923,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779736.652, "dur": 118.434,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779763.848, "dur": 88.041,
    "args": {
      "External id": 2488109,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657779772.858, "dur": 75.075,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657779871.771, "dur": 1.931,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514657779944.075, "dur": 1609.137,
    "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": 1336753, "tid": 1381210, "ts": 1514657779944.075,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1381210,
    "ts": 1514657780086.140, "dur": 101.209,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657780227.820, "dur": 38.867,
    "args": {
      "External id": 2488114,"kernel_hash": "c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/5t/c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr.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": 1336753, "tid": 1381210,
    "ts": 1514657780284.129, "dur": 46.247,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657780339.801, "dur": 31.720,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657780377.854, "dur": 42.782,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657780426.846, "dur": 26.648,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657780462.233, "dur": 40.730,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657780525.561, "dur": 39.656,
    "args": {
      "External id": 2488120,"kernel_hash": "cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/mm/cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k.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": 1336753, "tid": 1381210,
    "ts": 1514657780584.243, "dur": 29.035,
    "args": {
      "External id": 2488121,"kernel_hash": "c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2", "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/4b/c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2.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": 1336753, "tid": 1381210,
    "ts": 1514657780634.693, "dur": 21.583,
    "args": {
      "External id": 2488122,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657780668.883, "dur": 15.268,
    "args": {
      "External id": 2488123,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657780693.826, "dur": 34.140,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657780731.596, "dur": 31.597,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657780791.144, "dur": 162.668,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657780866.442, "dur": 5.979,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657780874.250, "dur": 2.692,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781022.271, "dur": 27.730,
    "args": {
      "External id": 2488129,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657781070.031, "dur": 15.941,
    "args": {
      "External id": 2488130,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657781095.318, "dur": 41.332,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781147.469, "dur": 33.081,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781187.089, "dur": 24.806,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781215.824, "dur": 28.416,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781249.507, "dur": 21.025,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781277.952, "dur": 28.471,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781325.236, "dur": 22.519,
    "args": {
      "External id": 2488137,"kernel_hash": "cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/as/cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf.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": 1336753, "tid": 1381210,
    "ts": 1514657781382.154, "dur": 32.371,
    "args": {
      "External id": 2488138,"kernel_hash": "cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od", "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/ny/cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od.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": 1336753, "tid": 1381210,
    "ts": 1514657781429.852, "dur": 17.784,
    "args": {
      "External id": 2488139,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657781463.625, "dur": 14.973,
    "args": {
      "External id": 2488140,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657781492.123, "dur": 16.282,
    "args": {
      "External id": 2488141,"kernel_hash": "c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/25/c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6.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": 1336753, "tid": 1381210,
    "ts": 1514657781596.704, "dur": 17.318,
    "args": {
      "External id": 2488142,"Record function id": 0, "Ev Idx": 4941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657781602.775, "dur": 10.271,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781606.590, "dur": 5.282,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781607.903, "dur": 3.857,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781617.828, "dur": 6.867,
    "args": {
      "External id": 2488146,"Record function id": 0, "Ev Idx": 4945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657781619.556, "dur": 4.692,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781620.218, "dur": 3.517,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781620.692, "dur": 2.985,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781627.962, "dur": 4.397,
    "args": {
      "External id": 2488150,"Record function id": 0, "Ev Idx": 4949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657781629.442, "dur": 2.479,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781629.935, "dur": 1.485,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781630.524, "dur": 0.800,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781635.533, "dur": 3.795,
    "args": {
      "External id": 2488154,"Record function id": 0, "Ev Idx": 4953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657781636.874, "dur": 2.025,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781637.463, "dur": 0.981,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781637.798, "dur": 0.556,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781642.462, "dur": 3.712,
    "args": {
      "External id": 2488158,"Record function id": 0, "Ev Idx": 4957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657781643.553, "dur": 2.215,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781644.099, "dur": 1.222,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781644.690, "dur": 0.554,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781649.535, "dur": 3.877,
    "args": {
      "External id": 2488162,"Record function id": 0, "Ev Idx": 4961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657781650.773, "dur": 2.230,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781651.232, "dur": 1.317,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781651.823, "dur": 0.657,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781656.567, "dur": 3.293,
    "args": {
      "External id": 2488166,"Record function id": 0, "Ev Idx": 4965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657781657.654, "dur": 1.812,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781658.138, "dur": 0.886,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781658.447, "dur": 0.513,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781662.926, "dur": 3.275,
    "args": {
      "External id": 2488170,"Record function id": 0, "Ev Idx": 4969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657781663.853, "dur": 1.959,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781664.286, "dur": 1.094,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781664.699, "dur": 0.614,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781669.334, "dur": 7.566,
    "args": {
      "External id": 2488174,"Record function id": 0, "Ev Idx": 4973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657781670.570, "dur": 5.917,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781670.993, "dur": 5.070,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781672.953, "dur": 2.998,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781680.499, "dur": 36754.477,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781681.712, "dur": 36745.410,
    "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": 1336753, "tid": 1381210, "ts": 1514657781681.712,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.5)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657781710.998, "dur": 35.166,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781757.040, "dur": 64.985,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781827.537, "dur": 36592.373,
    "args": {
      "External id": 2488182,"Record function id": 0, "Ev Idx": 4981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657781917.650, "dur": 6.851,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781937.006, "dur": 4.696,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781955.811, "dur": 35634.104,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657781968.903, "dur": 35612.504,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657782069.166, "dur": 14.371,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657782091.172, "dur": 35453.680,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657782093.963, "dur": 35450.270,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657782097.937, "dur": 9.520,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657782109.225, "dur": 35431.375,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657817673.542, "dur": 8.536,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657817676.560, "dur": 5.153,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657817712.545, "dur": 422.266,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657817741.142, "dur": 388.195,
    "args": {
      "External id": 2488195,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657817750.669, "dur": 372.661,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657818156.475, "dur": 2.588,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1381210,
    "ts": 1514657818220.486, "dur": 6.780,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657818268.738, "dur": 1.200,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657818288.648, "dur": 1.396,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657818303.546, "dur": 0.703,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657818315.042, "dur": 0.967,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657818325.634, "dur": 0.848,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657818339.381, "dur": 0.804,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657818351.018, "dur": 1.149,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657818361.740, "dur": 0.862,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657818448.589, "dur": 2652.980,
    "args": {
      "External id": 2488207,"Record function id": 0, "Ev Idx": 5006
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.4)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657818466.821, "dur": 999.589,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657818480.456, "dur": 319.095,
    "args": {
      "External id": 2488209,"Record function id": 0, "Ev Idx": 5008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1381210,
    "ts": 1514657818571.422, "dur": 4.594,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657818579.570, "dur": 0.887,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657818583.454, "dur": 2.793,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657818587.994, "dur": 0.849,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657818590.633, "dur": 0.713,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657818593.021, "dur": 0.955,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657818596.936, "dur": 0.906,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657818599.570, "dur": 0.662,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657818601.620, "dur": 0.927,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657818604.059, "dur": 0.933,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657818624.275, "dur": 147.199,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657818640.953, "dur": 126.300,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657818658.053, "dur": 13.624,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657818674.633, "dur": 64.289,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657818677.064, "dur": 61.458,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657818681.017, "dur": 5.816,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657818688.350, "dur": 49.543,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657818878.132, "dur": 580.876,
    "args": {
      "External id": 2488227,"Record function id": 0, "Ev Idx": 5026
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.3)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657818896.127, "dur": 550.503,
    "args": {
      "External id": 2488228,"Record function id": 0, "Ev Idx": 5027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657818952.040, "dur": 4.633,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657818975.336, "dur": 67.998,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657818980.856, "dur": 36.146,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657819021.729, "dur": 0.390,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657819023.884, "dur": 0.390,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657819025.567, "dur": 1.600,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657819028.848, "dur": 0.492,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657819030.803, "dur": 2.464,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657819036.010, "dur": 0.390,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657819037.673, "dur": 0.576,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657819040.006, "dur": 0.384,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657819055.546, "dur": 37.970,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657819125.250, "dur": 102.362,
    "args": {
      "External id": 2488241,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657819135.227, "dur": 4.048,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657819144.333, "dur": 10.327,
    "args": {
      "External id": 2488243,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657819148.149, "dur": 6.095,
    "args": {
      "External id": 2488244,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1381210,
    "ts": 1514657819151.731, "dur": 1.350,
    "args": {
      "External id": 2488245,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657819161.533, "dur": 26.240,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657819164.121, "dur": 0.387,
    "args": {
      "External id": 2488247,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657819165.893, "dur": 0.795,
    "args": {
      "External id": 2488248,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1381210,
    "ts": 1514657819168.272, "dur": 2.446,
    "args": {
      "External id": 2488249,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1381210,
    "ts": 1514657819172.641, "dur": 0.541,
    "args": {
      "External id": 2488250,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1381210,
    "ts": 1514657819174.711, "dur": 0.524,
    "args": {
      "External id": 2488251,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1381210,
    "ts": 1514657819176.655, "dur": 0.771,
    "args": {
      "External id": 2488252,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1381210,
    "ts": 1514657819178.724, "dur": 0.397,
    "args": {
      "External id": 2488253,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1381210,
    "ts": 1514657819180.226, "dur": 0.589,
    "args": {
      "External id": 2488254,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1381210,
    "ts": 1514657819182.521, "dur": 0.352,
    "args": {
      "External id": 2488255,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1381210,
    "ts": 1514657819199.417, "dur": 21.125,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657819272.008, "dur": 112.430,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657819299.234, "dur": 81.898,
    "args": {
      "External id": 2488258,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657819308.096, "dur": 69.121,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657819399.180, "dur": 2.071,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514657819473.106, "dur": 1609.441,
    "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": 1336753, "tid": 1381210, "ts": 1514657819473.106,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1381210,
    "ts": 1514657819595.533, "dur": 99.661,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657819736.760, "dur": 39.397,
    "args": {
      "External id": 2488263,"kernel_hash": "c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/5t/c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr.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": 1336753, "tid": 1381210,
    "ts": 1514657819791.147, "dur": 47.714,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657819848.256, "dur": 30.916,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657819886.805, "dur": 44.374,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657819937.043, "dur": 26.349,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657819969.338, "dur": 77.204,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657820074.144, "dur": 25.040,
    "args": {
      "External id": 2488269,"kernel_hash": "cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/mm/cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k.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": 1336753, "tid": 1381210,
    "ts": 1514657820118.514, "dur": 27.319,
    "args": {
      "External id": 2488270,"kernel_hash": "c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2", "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/4b/c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2.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": 1336753, "tid": 1381210,
    "ts": 1514657820165.823, "dur": 20.375,
    "args": {
      "External id": 2488271,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657820200.695, "dur": 16.272,
    "args": {
      "External id": 2488272,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657820225.322, "dur": 33.572,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657820262.037, "dur": 31.425,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657820321.051, "dur": 161.329,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657820394.249, "dur": 6.175,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657820402.145, "dur": 2.940,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657820513.932, "dur": 42.726,
    "args": {
      "External id": 2488278,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657820570.280, "dur": 17.421,
    "args": {
      "External id": 2488279,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657820597.035, "dur": 40.363,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657820643.035, "dur": 34.905,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657820683.738, "dur": 21.989,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657820712.418, "dur": 28.818,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657820746.580, "dur": 21.472,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657820773.765, "dur": 28.655,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657820821.382, "dur": 23.263,
    "args": {
      "External id": 2488286,"kernel_hash": "cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/as/cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf.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": 1336753, "tid": 1381210,
    "ts": 1514657820860.829, "dur": 23.556,
    "args": {
      "External id": 2488287,"kernel_hash": "cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od", "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/ny/cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od.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": 1336753, "tid": 1381210,
    "ts": 1514657820902.178, "dur": 17.745,
    "args": {
      "External id": 2488288,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657820935.273, "dur": 17.244,
    "args": {
      "External id": 2488289,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657820966.109, "dur": 77.476,
    "args": {
      "External id": 2488290,"kernel_hash": "c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/25/c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6.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": 1336753, "tid": 1381210,
    "ts": 1514657821124.651, "dur": 17.627,
    "args": {
      "External id": 2488291,"Record function id": 0, "Ev Idx": 5090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657821128.076, "dur": 13.348,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821135.393, "dur": 5.250,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821136.686, "dur": 3.856,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821146.134, "dur": 5.176,
    "args": {
      "External id": 2488295,"Record function id": 0, "Ev Idx": 5094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657821148.011, "dur": 2.839,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821148.914, "dur": 1.512,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821149.378, "dur": 0.975,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821154.648, "dur": 4.592,
    "args": {
      "External id": 2488299,"Record function id": 0, "Ev Idx": 5098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657821155.903, "dur": 2.913,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821156.555, "dur": 1.874,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821157.419, "dur": 0.937,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821162.470, "dur": 3.952,
    "args": {
      "External id": 2488303,"Record function id": 0, "Ev Idx": 5102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657821163.605, "dur": 2.410,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821164.352, "dur": 1.290,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821164.779, "dur": 0.788,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821169.645, "dur": 5.563,
    "args": {
      "External id": 2488307,"Record function id": 0, "Ev Idx": 5106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657821170.543, "dur": 4.254,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821171.413, "dur": 3.004,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821171.703, "dur": 2.639,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821178.377, "dur": 5.838,
    "args": {
      "External id": 2488311,"Record function id": 0, "Ev Idx": 5110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657821179.442, "dur": 4.346,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821179.984, "dur": 3.347,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821182.420, "dur": 0.843,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821190.632, "dur": 3.609,
    "args": {
      "External id": 2488315,"Record function id": 0, "Ev Idx": 5114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657821191.564, "dur": 2.251,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821192.173, "dur": 1.211,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821192.527, "dur": 0.784,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821197.371, "dur": 69.906,
    "args": {
      "External id": 2488319,"Record function id": 0, "Ev Idx": 5118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657821198.493, "dur": 68.219,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821264.626, "dur": 1.543,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821265.279, "dur": 0.805,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821270.693, "dur": 5.104,
    "args": {
      "External id": 2488323,"Record function id": 0, "Ev Idx": 5122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657821271.881, "dur": 3.457,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821272.426, "dur": 2.482,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821274.231, "dur": 0.566,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821279.627, "dur": 36173.028,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821281.496, "dur": 36163.214,
    "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": 1336753, "tid": 1381210, "ts": 1514657821281.496,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.4)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657821310.496, "dur": 37.975,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821356.273, "dur": 60.931,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821422.724, "dur": 36014.836,
    "args": {
      "External id": 2488331,"Record function id": 0, "Ev Idx": 5130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657821508.925, "dur": 7.030,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821524.615, "dur": 19.603,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821561.559, "dur": 35085.443,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821574.470, "dur": 35063.741,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821627.686, "dur": 13.479,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821649.049, "dur": 34950.293,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821651.488, "dur": 34947.196,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821655.835, "dur": 8.546,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657821666.033, "dur": 34929.231,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657856730.101, "dur": 8.696,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657856732.590, "dur": 5.859,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657856767.325, "dur": 387.560,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657856795.835, "dur": 353.995,
    "args": {
      "External id": 2488344,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657856805.358, "dur": 337.778,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657857178.781, "dur": 2.318,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1381210,
    "ts": 1514657857240.285, "dur": 6.354,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657857287.384, "dur": 1.302,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657857303.911, "dur": 1.362,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657857316.706, "dur": 2.685,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657857330.428, "dur": 0.734,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657857340.924, "dur": 0.715,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657857354.227, "dur": 0.932,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657857364.930, "dur": 3.044,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657857381.446, "dur": 0.811,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657857466.111, "dur": 2645.551,
    "args": {
      "External id": 2488356,"Record function id": 0, "Ev Idx": 5155
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.3)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657857484.121, "dur": 998.994,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657857497.852, "dur": 336.480,
    "args": {
      "External id": 2488358,"Record function id": 0, "Ev Idx": 5157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1381210,
    "ts": 1514657857589.820, "dur": 4.557,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657857597.878, "dur": 0.956,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657857601.961, "dur": 0.953,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657857604.622, "dur": 0.781,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657857606.721, "dur": 0.749,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657857608.859, "dur": 2.308,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657857613.497, "dur": 0.949,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657857616.219, "dur": 0.692,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657857618.900, "dur": 0.981,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657857621.607, "dur": 1.238,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657857641.780, "dur": 162.217,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657857657.718, "dur": 141.846,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657857693.714, "dur": 11.931,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657857708.832, "dur": 62.401,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657857711.368, "dur": 59.462,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657857715.010, "dur": 5.967,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657857722.548, "dur": 47.585,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657857915.084, "dur": 559.933,
    "args": {
      "External id": 2488376,"Record function id": 0, "Ev Idx": 5175
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.2)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657857931.876, "dur": 530.995,
    "args": {
      "External id": 2488377,"Record function id": 0, "Ev Idx": 5176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657858026.067, "dur": 6.321,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657858051.771, "dur": 28.168,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657858056.879, "dur": 4.335,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657858062.998, "dur": 0.419,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657858064.411, "dur": 0.373,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657858066.714, "dur": 0.320,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657858068.104, "dur": 0.310,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657858069.479, "dur": 0.376,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657858071.609, "dur": 0.326,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657858073.058, "dur": 0.428,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657858074.372, "dur": 2.432,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657858089.580, "dur": 34.078,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657858154.253, "dur": 91.234,
    "args": {
      "External id": 2488390,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657858163.556, "dur": 3.303,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657858171.626, "dur": 8.607,
    "args": {
      "External id": 2488392,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657858175.413, "dur": 4.407,
    "args": {
      "External id": 2488393,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1381210,
    "ts": 1514657858178.285, "dur": 0.398,
    "args": {
      "External id": 2488394,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657858186.504, "dur": 20.981,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657858188.479, "dur": 0.621,
    "args": {
      "External id": 2488396,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657858190.275, "dur": 0.585,
    "args": {
      "External id": 2488397,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1381210,
    "ts": 1514657858191.578, "dur": 1.068,
    "args": {
      "External id": 2488398,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1381210,
    "ts": 1514657858193.483, "dur": 0.431,
    "args": {
      "External id": 2488399,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1381210,
    "ts": 1514657858194.757, "dur": 0.340,
    "args": {
      "External id": 2488400,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1381210,
    "ts": 1514657858196.149, "dur": 2.338,
    "args": {
      "External id": 2488401,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1381210,
    "ts": 1514657858199.339, "dur": 0.433,
    "args": {
      "External id": 2488402,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1381210,
    "ts": 1514657858200.632, "dur": 0.452,
    "args": {
      "External id": 2488403,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1381210,
    "ts": 1514657858201.886, "dur": 0.621,
    "args": {
      "External id": 2488404,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1381210,
    "ts": 1514657858218.640, "dur": 19.710,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657858289.894, "dur": 110.020,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657858314.929, "dur": 81.544,
    "args": {
      "External id": 2488407,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657858323.831, "dur": 68.612,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657858414.168, "dur": 2.060,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514657858490.848, "dur": 1601.642,
    "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": 1336753, "tid": 1381210, "ts": 1514657858490.848,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1381210,
    "ts": 1514657858612.649, "dur": 99.560,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657858755.623, "dur": 40.554,
    "args": {
      "External id": 2488412,"kernel_hash": "c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/5t/c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr.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": 1336753, "tid": 1381210,
    "ts": 1514657858811.399, "dur": 45.534,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657858866.683, "dur": 31.641,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657858904.107, "dur": 43.979,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657858956.162, "dur": 62.129,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657859029.235, "dur": 45.478,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657859099.986, "dur": 24.360,
    "args": {
      "External id": 2488418,"kernel_hash": "cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/mm/cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k.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": 1336753, "tid": 1381210,
    "ts": 1514657859143.668, "dur": 27.220,
    "args": {
      "External id": 2488419,"kernel_hash": "c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2", "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/4b/c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2.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": 1336753, "tid": 1381210,
    "ts": 1514657859189.644, "dur": 20.041,
    "args": {
      "External id": 2488420,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657859224.376, "dur": 16.119,
    "args": {
      "External id": 2488421,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657859249.821, "dur": 29.948,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657859282.545, "dur": 31.367,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657859339.419, "dur": 162.290,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657859416.083, "dur": 5.763,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657859423.418, "dur": 1.988,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657859549.783, "dur": 26.640,
    "args": {
      "External id": 2488427,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657859588.325, "dur": 15.395,
    "args": {
      "External id": 2488428,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657859614.109, "dur": 38.114,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657859657.434, "dur": 32.452,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657859696.117, "dur": 21.478,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657859721.957, "dur": 28.846,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657859757.070, "dur": 20.798,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657859783.231, "dur": 28.607,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657859829.839, "dur": 22.625,
    "args": {
      "External id": 2488435,"kernel_hash": "cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/as/cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf.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": 1336753, "tid": 1381210,
    "ts": 1514657859869.139, "dur": 24.435,
    "args": {
      "External id": 2488436,"kernel_hash": "cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od", "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/ny/cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od.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": 1336753, "tid": 1381210,
    "ts": 1514657859907.030, "dur": 30.980,
    "args": {
      "External id": 2488437,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657859962.956, "dur": 16.962,
    "args": {
      "External id": 2488438,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657860042.380, "dur": 19.274,
    "args": {
      "External id": 2488439,"kernel_hash": "c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/25/c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6.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": 1336753, "tid": 1381210,
    "ts": 1514657860136.116, "dur": 15.784,
    "args": {
      "External id": 2488440,"Record function id": 0, "Ev Idx": 5239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657860139.307, "dur": 11.705,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860143.404, "dur": 6.712,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860144.769, "dur": 5.263,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860155.828, "dur": 4.448,
    "args": {
      "External id": 2488444,"Record function id": 0, "Ev Idx": 5243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657860157.292, "dur": 2.526,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860158.095, "dur": 1.296,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860158.529, "dur": 0.768,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860163.503, "dur": 4.074,
    "args": {
      "External id": 2488448,"Record function id": 0, "Ev Idx": 5247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657860164.676, "dur": 2.474,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860165.170, "dur": 1.601,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860165.855, "dur": 0.800,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860170.842, "dur": 3.781,
    "args": {
      "External id": 2488452,"Record function id": 0, "Ev Idx": 5251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657860171.841, "dur": 2.343,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860172.489, "dur": 1.293,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860172.969, "dur": 0.690,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860177.855, "dur": 3.680,
    "args": {
      "External id": 2488456,"Record function id": 0, "Ev Idx": 5255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657860179.123, "dur": 1.990,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860179.733, "dur": 0.993,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860180.001, "dur": 0.631,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860184.744, "dur": 3.517,
    "args": {
      "External id": 2488460,"Record function id": 0, "Ev Idx": 5259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657860185.727, "dur": 2.114,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860186.217, "dur": 1.198,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860186.766, "dur": 0.555,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860191.533, "dur": 4.937,
    "args": {
      "External id": 2488464,"Record function id": 0, "Ev Idx": 5263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657860192.564, "dur": 3.492,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860193.183, "dur": 2.471,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860194.992, "dur": 0.547,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860199.627, "dur": 10.342,
    "args": {
      "External id": 2488468,"Record function id": 0, "Ev Idx": 5267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657860200.933, "dur": 8.586,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860205.995, "dur": 3.111,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860206.626, "dur": 2.412,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860213.067, "dur": 3.672,
    "args": {
      "External id": 2488472,"Record function id": 0, "Ev Idx": 5271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657860214.321, "dur": 1.989,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860214.809, "dur": 1.123,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860215.107, "dur": 0.731,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860220.866, "dur": 38478.313,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860222.221, "dur": 38468.660,
    "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": 1336753, "tid": 1381210, "ts": 1514657860222.221,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.3)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657860248.297, "dur": 32.296,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860287.894, "dur": 57.826,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860350.819, "dur": 38332.434,
    "args": {
      "External id": 2488480,"Record function id": 0, "Ev Idx": 5279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657860430.572, "dur": 6.630,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860445.768, "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": 1336753, "tid": 1381210,
    "ts": 1514657860464.950, "dur": 37361.839,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860477.066, "dur": 37341.050,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860551.417, "dur": 13.909,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860571.783, "dur": 37207.024,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860574.365, "dur": 37203.789,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860578.105, "dur": 5.977,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657860585.591, "dur": 37189.154,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657897907.911, "dur": 8.282,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657897910.268, "dur": 5.550,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657897945.556, "dur": 434.392,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657897972.773, "dur": 401.960,
    "args": {
      "External id": 2488493,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657898018.299, "dur": 350.183,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657898400.694, "dur": 2.067,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1381210,
    "ts": 1514657898462.847, "dur": 6.571,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657898511.321, "dur": 1.244,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657898542.478, "dur": 3.865,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657898562.565, "dur": 0.724,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657898575.001, "dur": 0.657,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657898585.694, "dur": 0.937,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657898597.061, "dur": 2.494,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657898613.263, "dur": 1.033,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657898623.696, "dur": 0.864,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657898714.135, "dur": 2708.503,
    "args": {
      "External id": 2488505,"Record function id": 0, "Ev Idx": 5304
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.2)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657898733.697, "dur": 1032.882,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657898747.475, "dur": 343.533,
    "args": {
      "External id": 2488507,"Record function id": 0, "Ev Idx": 5306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1381210,
    "ts": 1514657898827.021, "dur": 3.889,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657898833.676, "dur": 1.252,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657898836.560, "dur": 0.856,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657898840.385, "dur": 2.475,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657898844.236, "dur": 0.793,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657898846.383, "dur": 0.677,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657898849.018, "dur": 0.839,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657898852.129, "dur": 0.753,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657898854.135, "dur": 1.050,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657898856.527, "dur": 1.377,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657898875.850, "dur": 184.733,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657898890.907, "dur": 164.643,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657898907.989, "dur": 12.342,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657898923.338, "dur": 102.443,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657898927.438, "dur": 97.922,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657898931.109, "dur": 8.167,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657898940.886, "dur": 83.514,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657899176.822, "dur": 581.114,
    "args": {
      "External id": 2488525,"Record function id": 0, "Ev Idx": 5324
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.1)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657899196.817, "dur": 549.028,
    "args": {
      "External id": 2488526,"Record function id": 0, "Ev Idx": 5325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657899260.641, "dur": 5.267,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657899312.489, "dur": 26.394,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657899317.533, "dur": 1.718,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657899321.697, "dur": 0.378,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657899323.033, "dur": 0.521,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657899325.155, "dur": 0.425,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657899327.475, "dur": 0.690,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657899329.087, "dur": 0.516,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657899330.650, "dur": 2.384,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657899334.268, "dur": 0.266,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657899335.542, "dur": 0.331,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657899350.391, "dur": 32.155,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657899414.180, "dur": 96.539,
    "args": {
      "External id": 2488539,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657899423.788, "dur": 2.763,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657899431.592, "dur": 9.798,
    "args": {
      "External id": 2488541,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657899435.715, "dur": 5.268,
    "args": {
      "External id": 2488542,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1381210,
    "ts": 1514657899439.096, "dur": 0.779,
    "args": {
      "External id": 2488543,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657899448.455, "dur": 26.410,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657899450.149, "dur": 0.650,
    "args": {
      "External id": 2488545,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657899452.316, "dur": 0.410,
    "args": {
      "External id": 2488546,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1381210,
    "ts": 1514657899453.845, "dur": 0.374,
    "args": {
      "External id": 2488547,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1381210,
    "ts": 1514657899455.528, "dur": 1.870,
    "args": {
      "External id": 2488548,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1381210,
    "ts": 1514657899459.295, "dur": 0.366,
    "args": {
      "External id": 2488549,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1381210,
    "ts": 1514657899460.981, "dur": 0.348,
    "args": {
      "External id": 2488550,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1381210,
    "ts": 1514657899465.166, "dur": 0.353,
    "args": {
      "External id": 2488551,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1381210,
    "ts": 1514657899466.680, "dur": 0.333,
    "args": {
      "External id": 2488552,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1381210,
    "ts": 1514657899468.443, "dur": 0.358,
    "args": {
      "External id": 2488553,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1381210,
    "ts": 1514657899484.441, "dur": 19.145,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657899571.054, "dur": 110.319,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657899594.658, "dur": 83.489,
    "args": {
      "External id": 2488556,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657899603.646, "dur": 70.528,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657899696.124, "dur": 1.633,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514657899773.205, "dur": 1629.979,
    "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": 1336753, "tid": 1381210, "ts": 1514657899773.205,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1381210,
    "ts": 1514657899878.804, "dur": 99.246,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657900057.778, "dur": 46.786,
    "args": {
      "External id": 2488561,"kernel_hash": "c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/5t/c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr.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": 1336753, "tid": 1381210,
    "ts": 1514657900121.648, "dur": 54.719,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657900185.865, "dur": 32.580,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657900224.847, "dur": 43.519,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657900276.309, "dur": 28.135,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657900314.222, "dur": 41.947,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657900379.902, "dur": 23.960,
    "args": {
      "External id": 2488567,"kernel_hash": "cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/mm/cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k.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": 1336753, "tid": 1381210,
    "ts": 1514657900421.400, "dur": 29.487,
    "args": {
      "External id": 2488568,"kernel_hash": "c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2", "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/4b/c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2.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": 1336753, "tid": 1381210,
    "ts": 1514657900471.868, "dur": 19.790,
    "args": {
      "External id": 2488569,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657900505.038, "dur": 15.891,
    "args": {
      "External id": 2488570,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657900547.096, "dur": 37.069,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657900588.213, "dur": 34.708,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657900653.720, "dur": 167.966,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657900731.719, "dur": 6.071,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657900739.871, "dur": 2.182,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657900853.132, "dur": 26.068,
    "args": {
      "External id": 2488576,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657900890.799, "dur": 15.460,
    "args": {
      "External id": 2488577,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657900915.782, "dur": 34.171,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657900957.255, "dur": 74.697,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901043.044, "dur": 27.075,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901075.987, "dur": 30.391,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901114.262, "dur": 22.276,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901143.727, "dur": 42.155,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901210.041, "dur": 29.921,
    "args": {
      "External id": 2488584,"kernel_hash": "cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/as/cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf.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": 1336753, "tid": 1381210,
    "ts": 1514657901256.317, "dur": 27.187,
    "args": {
      "External id": 2488585,"kernel_hash": "cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od", "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/ny/cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od.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": 1336753, "tid": 1381210,
    "ts": 1514657901297.537, "dur": 17.321,
    "args": {
      "External id": 2488586,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657901327.243, "dur": 19.204,
    "args": {
      "External id": 2488587,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657901357.940, "dur": 15.487,
    "args": {
      "External id": 2488588,"kernel_hash": "c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/25/c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6.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": 1336753, "tid": 1381210,
    "ts": 1514657901444.308, "dur": 17.359,
    "args": {
      "External id": 2488589,"Record function id": 0, "Ev Idx": 5388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657901447.489, "dur": 13.428,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901451.379, "dur": 8.791,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901452.875, "dur": 7.147,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901465.737, "dur": 4.971,
    "args": {
      "External id": 2488593,"Record function id": 0, "Ev Idx": 5392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657901466.936, "dur": 3.276,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901467.924, "dur": 1.806,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901468.718, "dur": 0.923,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901473.991, "dur": 4.126,
    "args": {
      "External id": 2488597,"Record function id": 0, "Ev Idx": 5396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657901475.150, "dur": 2.559,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901475.707, "dur": 1.579,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901476.196, "dur": 1.019,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901481.451, "dur": 6.546,
    "args": {
      "External id": 2488601,"Record function id": 0, "Ev Idx": 5400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657901482.723, "dur": 4.854,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901483.559, "dur": 3.609,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901484.146, "dur": 2.953,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901491.171, "dur": 3.940,
    "args": {
      "External id": 2488605,"Record function id": 0, "Ev Idx": 5404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657901492.595, "dur": 2.088,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901493.049, "dur": 1.246,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901493.382, "dur": 0.844,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901498.394, "dur": 3.359,
    "args": {
      "External id": 2488609,"Record function id": 0, "Ev Idx": 5408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657901499.359, "dur": 1.936,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901499.836, "dur": 1.032,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901500.261, "dur": 0.498,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901505.264, "dur": 4.489,
    "args": {
      "External id": 2488613,"Record function id": 0, "Ev Idx": 5412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657901506.623, "dur": 2.693,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901507.513, "dur": 1.408,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901508.218, "dur": 0.641,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901516.068, "dur": 3.847,
    "args": {
      "External id": 2488617,"Record function id": 0, "Ev Idx": 5416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657901517.383, "dur": 2.078,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901518.006, "dur": 1.081,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901518.296, "dur": 0.678,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901523.018, "dur": 3.621,
    "args": {
      "External id": 2488621,"Record function id": 0, "Ev Idx": 5420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657901523.967, "dur": 2.236,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901524.591, "dur": 1.232,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901525.080, "dur": 0.624,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901548.086, "dur": 36335.872,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901550.082, "dur": 36326.093,
    "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": 1336753, "tid": 1381210, "ts": 1514657901550.082,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.2)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657901578.792, "dur": 36.018,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901622.949, "dur": 70.070,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901698.796, "dur": 36170.348,
    "args": {
      "External id": 2488629,"Record function id": 0, "Ev Idx": 5428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657901788.354, "dur": 6.400,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901803.643, "dur": 4.688,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901822.713, "dur": 35271.078,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901836.055, "dur": 35249.046,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901898.105, "dur": 13.126,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901917.239, "dur": 35129.092,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901919.693, "dur": 35125.887,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901924.282, "dur": 10.610,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657901937.723, "dur": 35104.362,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657937178.586, "dur": 9.582,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657937181.421, "dur": 6.412,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657937216.023, "dur": 374.138,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657937243.694, "dur": 341.338,
    "args": {
      "External id": 2488642,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657937253.957, "dur": 324.681,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657937610.458, "dur": 2.114,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1381210,
    "ts": 1514657937670.654, "dur": 6.568,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657937720.258, "dur": 1.528,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657937737.254, "dur": 1.156,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657937749.927, "dur": 2.484,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657937763.561, "dur": 0.930,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657937774.740, "dur": 0.982,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657937785.961, "dur": 1.130,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657937797.655, "dur": 3.440,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657937811.198, "dur": 1.086,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657937897.228, "dur": 2684.573,
    "args": {
      "External id": 2488654,"Record function id": 0, "Ev Idx": 5453
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.1)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657937915.790, "dur": 999.658,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657937929.177, "dur": 343.080,
    "args": {
      "External id": 2488656,"Record function id": 0, "Ev Idx": 5455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1381210,
    "ts": 1514657938046.715, "dur": 4.943,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938055.334, "dur": 0.709,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938057.505, "dur": 0.945,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938060.285, "dur": 0.853,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938062.536, "dur": 1.053,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938064.876, "dur": 0.840,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938068.498, "dur": 2.998,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938072.983, "dur": 0.836,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938075.175, "dur": 1.075,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938077.909, "dur": 1.367,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938098.925, "dur": 144.545,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938114.459, "dur": 125.067,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938131.454, "dur": 11.945,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938146.076, "dur": 64.867,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938148.590, "dur": 61.945,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938152.407, "dur": 6.603,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938160.865, "dur": 48.877,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938366.930, "dur": 540.438,
    "args": {
      "External id": 2488674,"Record function id": 0, "Ev Idx": 5473
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.0)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657938386.993, "dur": 508.471,
    "args": {
      "External id": 2488675,"Record function id": 0, "Ev Idx": 5474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657938451.191, "dur": 4.319,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938471.442, "dur": 25.672,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938476.002, "dur": 1.628,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938479.774, "dur": 0.691,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938481.887, "dur": 2.201,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938485.384, "dur": 0.652,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938487.275, "dur": 0.341,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938488.962, "dur": 0.459,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938490.713, "dur": 0.419,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938492.327, "dur": 0.583,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938494.172, "dur": 0.568,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938507.244, "dur": 48.479,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938588.475, "dur": 97.573,
    "args": {
      "External id": 2488688,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657938598.606, "dur": 4.114,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938607.827, "dur": 12.097,
    "args": {
      "External id": 2488690,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1381210,
    "ts": 1514657938611.793, "dur": 7.696,
    "args": {
      "External id": 2488691,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1381210,
    "ts": 1514657938615.335, "dur": 2.455,
    "args": {
      "External id": 2488692,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657938626.190, "dur": 21.956,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938628.176, "dur": 0.531,
    "args": {
      "External id": 2488694,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1381210,
    "ts": 1514657938630.192, "dur": 0.352,
    "args": {
      "External id": 2488695,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1381210,
    "ts": 1514657938631.730, "dur": 0.338,
    "args": {
      "External id": 2488696,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1381210,
    "ts": 1514657938633.201, "dur": 0.446,
    "args": {
      "External id": 2488697,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1381210,
    "ts": 1514657938634.953, "dur": 0.506,
    "args": {
      "External id": 2488698,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1381210,
    "ts": 1514657938636.746, "dur": 0.351,
    "args": {
      "External id": 2488699,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1381210,
    "ts": 1514657938638.253, "dur": 0.353,
    "args": {
      "External id": 2488700,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1381210,
    "ts": 1514657938639.820, "dur": 2.045,
    "args": {
      "External id": 2488701,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1381210,
    "ts": 1514657938642.741, "dur": 0.434,
    "args": {
      "External id": 2488702,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1381210,
    "ts": 1514657938658.098, "dur": 20.744,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938729.432, "dur": 103.358,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938751.576, "dur": 78.076,
    "args": {
      "External id": 2488705,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657938760.633, "dur": 65.154,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657938848.076, "dur": 1.836,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336753, "tid": 1381210,
    "ts": 1514657938922.216, "dur": 1637.532,
    "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": 1336753, "tid": 1381210, "ts": 1514657938922.216,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1381210,
    "ts": 1514657939071.704, "dur": 101.558,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657939213.277, "dur": 39.100,
    "args": {
      "External id": 2488710,"kernel_hash": "c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/5t/c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr.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": 1336753, "tid": 1381210,
    "ts": 1514657939268.152, "dur": 45.345,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657939324.977, "dur": 30.928,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657939362.574, "dur": 43.292,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657939412.454, "dur": 26.345,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657939445.197, "dur": 40.747,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657939509.894, "dur": 39.252,
    "args": {
      "External id": 2488716,"kernel_hash": "cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/mm/cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k.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": 1336753, "tid": 1381210,
    "ts": 1514657939575.587, "dur": 30.012,
    "args": {
      "External id": 2488717,"kernel_hash": "c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2", "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/4b/c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2.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": 1336753, "tid": 1381210,
    "ts": 1514657939627.705, "dur": 23.892,
    "args": {
      "External id": 2488718,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657939665.620, "dur": 17.109,
    "args": {
      "External id": 2488719,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657939691.468, "dur": 34.158,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657939728.783, "dur": 34.458,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657939790.550, "dur": 163.294,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657939865.665, "dur": 5.481,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657939873.373, "dur": 1.766,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940017.966, "dur": 29.213,
    "args": {
      "External id": 2488725,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657940059.510, "dur": 15.306,
    "args": {
      "External id": 2488726,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657940085.193, "dur": 43.859,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940135.709, "dur": 36.321,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940179.804, "dur": 22.506,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940207.406, "dur": 32.614,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940245.825, "dur": 20.442,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940273.623, "dur": 32.142,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940321.509, "dur": 24.955,
    "args": {
      "External id": 2488733,"kernel_hash": "cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/as/cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf.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": 1336753, "tid": 1381210,
    "ts": 1514657940371.819, "dur": 38.008,
    "args": {
      "External id": 2488734,"kernel_hash": "cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od", "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/ny/cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od.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": 1336753, "tid": 1381210,
    "ts": 1514657940434.992, "dur": 19.350,
    "args": {
      "External id": 2488735,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657940469.470, "dur": 18.499,
    "args": {
      "External id": 2488736,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657940500.752, "dur": 15.659,
    "args": {
      "External id": 2488737,"kernel_hash": "c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/25/c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6.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": 1336753, "tid": 1381210,
    "ts": 1514657940607.109, "dur": 16.950,
    "args": {
      "External id": 2488738,"Record function id": 0, "Ev Idx": 5537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657940610.448, "dur": 12.669,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940614.428, "dur": 7.762,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940616.533, "dur": 5.545,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940627.884, "dur": 5.438,
    "args": {
      "External id": 2488742,"Record function id": 0, "Ev Idx": 5541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657940629.522, "dur": 3.327,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940630.429, "dur": 1.908,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940631.249, "dur": 1.023,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940636.674, "dur": 3.988,
    "args": {
      "External id": 2488746,"Record function id": 0, "Ev Idx": 5545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657940638.260, "dur": 1.968,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940638.855, "dur": 0.954,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940639.175, "dur": 0.562,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940643.874, "dur": 6.637,
    "args": {
      "External id": 2488750,"Record function id": 0, "Ev Idx": 5549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657940647.975, "dur": 2.130,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940648.471, "dur": 1.216,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940648.961, "dur": 0.664,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940653.642, "dur": 3.676,
    "args": {
      "External id": 2488754,"Record function id": 0, "Ev Idx": 5553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657940654.995, "dur": 1.907,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940655.454, "dur": 1.037,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940655.768, "dur": 0.646,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940660.452, "dur": 3.838,
    "args": {
      "External id": 2488758,"Record function id": 0, "Ev Idx": 5557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657940661.588, "dur": 2.258,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940662.100, "dur": 1.338,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940662.547, "dur": 0.760,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940667.653, "dur": 3.962,
    "args": {
      "External id": 2488762,"Record function id": 0, "Ev Idx": 5561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657940669.153, "dur": 2.044,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940669.650, "dur": 1.149,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940669.919, "dur": 0.798,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940674.749, "dur": 5.004,
    "args": {
      "External id": 2488766,"Record function id": 0, "Ev Idx": 5565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657940676.100, "dur": 3.200,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940676.562, "dur": 2.345,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940676.840, "dur": 1.964,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940682.897, "dur": 3.976,
    "args": {
      "External id": 2488770,"Record function id": 0, "Ev Idx": 5569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657940684.130, "dur": 2.296,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940684.579, "dur": 1.445,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940685.042, "dur": 0.873,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940690.873, "dur": 36938.576,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940692.006, "dur": 36928.814,
    "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": 1336753, "tid": 1381210, "ts": 1514657940692.006,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.1)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657940719.171, "dur": 36.271,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940763.625, "dur": 67.088,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940840.052, "dur": 36773.429,
    "args": {
      "External id": 2488778,"Record function id": 0, "Ev Idx": 5577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657940923.448, "dur": 6.367,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940938.559, "dur": 4.682,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940956.781, "dur": 35869.580,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657940969.159, "dur": 35849.072,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657941042.542, "dur": 13.895,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657941062.806, "dur": 35717.975,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657941065.494, "dur": 35714.639,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657941069.484, "dur": 6.130,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657941077.366, "dur": 35699.596,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657976910.292, "dur": 9.617,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657976914.572, "dur": 4.977,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657976947.607, "dur": 371.259,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657976972.851, "dur": 341.579,
    "args": {
      "External id": 2488791,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514657977018.531, "dur": 290.934,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657977337.175, "dur": 2.280,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1381210,
    "ts": 1514657977396.777, "dur": 6.512,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657977445.737, "dur": 0.997,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657977462.720, "dur": 2.549,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657977477.453, "dur": 0.889,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657977489.274, "dur": 1.005,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657977498.984, "dur": 1.016,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657977510.642, "dur": 2.319,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657977522.891, "dur": 0.959,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657977552.233, "dur": 1.327,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657977644.271, "dur": 2172.248,
    "args": {
      "External id": 2488803,"Record function id": 0, "Ev Idx": 5602
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.0)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657977663.434, "dur": 478.019,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657977678.452, "dur": 348.438,
    "args": {
      "External id": 2488805,"Record function id": 0, "Ev Idx": 5604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1381210,
    "ts": 1514657977768.712, "dur": 3.835,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657977775.627, "dur": 0.920,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657977778.085, "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": 1336753, "tid": 1381210,
    "ts": 1514657977780.806, "dur": 0.926,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657977783.406, "dur": 2.540,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657977788.105, "dur": 1.035,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657977790.690, "dur": 1.007,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657977793.259, "dur": 0.807,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657977795.572, "dur": 1.099,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657977798.713, "dur": 0.753,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657977816.031, "dur": 143.486,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657977833.069, "dur": 122.352,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657977848.936, "dur": 11.270,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657977863.266, "dur": 63.564,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657977865.643, "dur": 60.862,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657977869.813, "dur": 5.759,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657977877.149, "dur": 48.570,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657978149.413, "dur": 1646.970,
    "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": 1336753, "tid": 1381210, "ts": 1514657978149.413,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1381210,
    "ts": 1514657978255.334, "dur": 98.806,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657978393.003, "dur": 42.251,
    "args": {
      "External id": 2488825,"kernel_hash": "c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/5t/c5tnhs5xq5qc7yr3l7cgs4vzxtk3cliycxpzqbsxievrlofzpfhr.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": 1336753, "tid": 1381210,
    "ts": 1514657978452.296, "dur": 47.633,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657978508.494, "dur": 50.066,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657978569.948, "dur": 51.382,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657978628.730, "dur": 28.049,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657978663.623, "dur": 42.455,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657978728.536, "dur": 23.794,
    "args": {
      "External id": 2488831,"kernel_hash": "cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/mm/cmmtfvdkgvqhsz243szg36eql3ea4ycfjtkqpfgjhi44lbeji46k.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": 1336753, "tid": 1381210,
    "ts": 1514657978768.566, "dur": 29.865,
    "args": {
      "External id": 2488832,"kernel_hash": "c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2", "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/4b/c4bvpuxlruf4ls6zaebh4ulq5kwzna7xpak4xntuufnkxt6y4wc2.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": 1336753, "tid": 1381210,
    "ts": 1514657978885.718, "dur": 24.101,
    "args": {
      "External id": 2488833,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657978924.276, "dur": 16.114,
    "args": {
      "External id": 2488834,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657978950.277, "dur": 75.925,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979032.345, "dur": 39.343,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979100.382, "dur": 173.651,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979178.918, "dur": 5.364,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979186.112, "dur": 3.215,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979306.705, "dur": 26.683,
    "args": {
      "External id": 2488840,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657979343.772, "dur": 16.177,
    "args": {
      "External id": 2488841,"kernel_hash": "cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t", "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/d3/cd3ye4lv3yv2y7l5dde5joralcg7ll6hlrrxvigczijvlwrp2l6t.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": 1336753, "tid": 1381210,
    "ts": 1514657979368.102, "dur": 35.265,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979409.803, "dur": 34.555,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979450.637, "dur": 21.980,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979479.218, "dur": 29.880,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979515.054, "dur": 37.244,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979561.813, "dur": 33.625,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979611.706, "dur": 27.405,
    "args": {
      "External id": 2488848,"kernel_hash": "cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/as/cas6dq7e5rikp56ssrhbxwcqp4ttdyuwqorivviwplq7rgsqykgf.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": 1336753, "tid": 1381210,
    "ts": 1514657979654.568, "dur": 25.265,
    "args": {
      "External id": 2488849,"kernel_hash": "cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od", "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/ny/cnyczubjxqoksn4kug4drlj6rdh2oloxs37rovzz3o7ct4qcv4od.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": 1336753, "tid": 1381210,
    "ts": 1514657979695.986, "dur": 17.641,
    "args": {
      "External id": 2488850,"kernel_hash": "c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/2k/c2kpn4mfmx5wt5gmjhm7dpm4gulqcmz5mcm7dlheptufg56to5av.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": 1336753, "tid": 1381210,
    "ts": 1514657979726.487, "dur": 16.762,
    "args": {
      "External id": 2488851,"kernel_hash": "cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/yf/cyfeqwmtr3blvks63uqsw3zds2osxyp6k5wrjhmcluio7cothlaw.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": 1336753, "tid": 1381210,
    "ts": 1514657979754.147, "dur": 15.833,
    "args": {
      "External id": 2488852,"kernel_hash": "c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/25/c25b4nhnmm2ykwmscwd3eqlx5bkfkwuzdtym5ifpvjo2smy2yhw6.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": 1336753, "tid": 1381210,
    "ts": 1514657979841.647, "dur": 15.508,
    "args": {
      "External id": 2488853,"Record function id": 0, "Ev Idx": 5652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657979844.714, "dur": 11.583,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979848.432, "dur": 6.901,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979849.912, "dur": 5.306,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979861.220, "dur": 4.689,
    "args": {
      "External id": 2488857,"Record function id": 0, "Ev Idx": 5656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657979862.724, "dur": 2.752,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979863.331, "dur": 1.680,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979864.019, "dur": 0.921,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979869.161, "dur": 4.643,
    "args": {
      "External id": 2488861,"Record function id": 0, "Ev Idx": 5660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657979870.838, "dur": 2.579,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979871.662, "dur": 1.337,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979872.053, "dur": 0.878,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979876.954, "dur": 3.956,
    "args": {
      "External id": 2488865,"Record function id": 0, "Ev Idx": 5664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657979878.449, "dur": 2.082,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979879.079, "dur": 1.008,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979879.376, "dur": 0.649,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979883.996, "dur": 6.990,
    "args": {
      "External id": 2488869,"Record function id": 0, "Ev Idx": 5668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657979888.742, "dur": 1.862,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979889.238, "dur": 0.971,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979889.551, "dur": 0.590,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979894.162, "dur": 4.359,
    "args": {
      "External id": 2488873,"Record function id": 0, "Ev Idx": 5672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657979895.637, "dur": 2.473,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979896.307, "dur": 1.361,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979896.767, "dur": 0.791,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979901.849, "dur": 4.318,
    "args": {
      "External id": 2488877,"Record function id": 0, "Ev Idx": 5676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657979902.922, "dur": 2.833,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979903.740, "dur": 1.567,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979904.388, "dur": 0.831,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979909.309, "dur": 4.835,
    "args": {
      "External id": 2488881,"Record function id": 0, "Ev Idx": 5680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657979910.609, "dur": 3.104,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979911.229, "dur": 2.100,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979911.497, "dur": 1.752,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979917.289, "dur": 3.946,
    "args": {
      "External id": 2488885,"Record function id": 0, "Ev Idx": 5684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514657979918.384, "dur": 2.416,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979919.028, "dur": 1.354,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979919.517, "dur": 0.751,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979924.840, "dur": 36666.956,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657979925.943, "dur": 36657.125,
    "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": 1336753, "tid": 1381210, "ts": 1514657979925.943,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.0)", "pid": 1336753, "tid": 1381210,
    "ts": 1514657979954.506, "dur": 76.128,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657980039.591, "dur": 85.107,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657980135.831, "dur": 36438.442,
    "args": {
      "External id": 2488893,"Record function id": 0, "Ev Idx": 5692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514657980228.883, "dur": 6.722,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657980245.767, "dur": 5.166,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657980265.337, "dur": 35493.643,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657980277.868, "dur": 35472.257,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657980333.481, "dur": 17.178,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657980357.181, "dur": 35355.867,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657980359.507, "dur": 35352.808,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657980363.872, "dur": 5.063,
    "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": 1336753, "tid": 1381210,
    "ts": 1514657980370.622, "dur": 35338.521,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658015844.653, "dur": 8.690,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658015847.277, "dur": 5.754,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658015883.427, "dur": 394.540,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658015908.493, "dur": 364.508,
    "args": {
      "External id": 2488906,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514658015918.270, "dur": 349.136,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658016298.929, "dur": 2.663,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1381210,
    "ts": 1514658016360.738, "dur": 6.428,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658016410.534, "dur": 1.465,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658016427.248, "dur": 2.712,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658016441.906, "dur": 0.851,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658016455.278, "dur": 0.867,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658016465.897, "dur": 1.053,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658016476.899, "dur": 1.902,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658016489.940, "dur": 1.084,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658016500.394, "dur": 0.711,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658016607.527, "dur": 276.244,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658016610.100, "dur": 265.955,
    "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": 1336753, "tid": 1381210, "ts": 1514658016610.100,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_embedding_dense_backward_0", "pid": 1336753, "tid": 1381210,
    "ts": 1514658016726.183, "dur": 48.767,
    "args": {
      "External id": 2488920,"kernel_hash": "cfr73mqqwqpdrss7r6csecalgrfowe5r3r24yaeyivzpezfoy4fz", "grid": "grid(65536000,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "65536000"], "kernel_file": "/tmp/torchinductor_cvm/fr/cfr73mqqwqpdrss7r6csecalgrfowe5r3r24yaeyivzpezfoy4fz.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": 1336753, "tid": 1381210,
    "ts": 1514658016791.975, "dur": 29.510,
    "args": {
      "External id": 2488921,"kernel_hash": "cxedqnbrx2bvln4bf3pphdmthxjm6nsxqowbdddlxye4bt3yfnsg", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/xe/cxedqnbrx2bvln4bf3pphdmthxjm6nsxqowbdddlxye4bt3yfnsg.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": 1336753, "tid": 1381210,
    "ts": 1514658016838.279, "dur": 23.400,
    "args": {
      "External id": 2488922,"kernel_hash": "c2xxuncnwkfhjvqk7hfnejnsljjkm6qo4ajwnumdr7vdpf4nk2iq", "grid": "grid(65536000,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "65536000"], "kernel_file": "/tmp/torchinductor_cvm/2x/c2xxuncnwkfhjvqk7hfnejnsljjkm6qo4ajwnumdr7vdpf4nk2iq.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": 1336753, "tid": 1381210,
    "ts": 1514658016892.835, "dur": 14.168,
    "args": {
      "External id": 2488923,"Record function id": 0, "Ev Idx": 5722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336753, "tid": 1381210,
    "ts": 1514658016896.208, "dur": 10.095,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658016899.861, "dur": 5.635,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658016901.523, "dur": 3.880,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658016928.292, "dur": 9905.116,
    "args": {
      "External id": 2488927,"Record function id": 0, "Ev Idx": 5726
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate", "pid": 1336753, "tid": 1381210,
    "ts": 1514658016946.223, "dur": 85.471,
    "args": {
      "External id": 2488928,"Record function id": 0, "Ev Idx": 5727
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard", "pid": 1336753, "tid": 1381210,
    "ts": 1514658017040.721, "dur": 259.362,
    "args": {
      "External id": 2488929,"Record function id": 0, "Ev Idx": 5728
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce", "pid": 1336753, "tid": 1381210,
    "ts": 1514658017306.567, "dur": 9258.364,
    "args": {
      "External id": 2488930,"Record function id": 0, "Ev Idx": 5729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1381210,
    "ts": 1514658017427.895, "dur": 7.167,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658017445.593, "dur": 5.187,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658017471.463, "dur": 7865.786,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658017489.681, "dur": 7835.304,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658017823.834, "dur": 18.648,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658017934.562, "dur": 7338.403,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658017937.712, "dur": 7334.296,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658017942.545, "dur": 13.291,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658017957.633, "dur": 7309.481,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658025466.484, "dur": 11.680,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658025470.388, "dur": 7.258,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658025505.288, "dur": 382.678,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658025542.637, "dur": 340.709,
    "args": {
      "External id": 2488943,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1381210,
    "ts": 1514658025558.631, "dur": 319.137,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658025905.296, "dur": 2.373,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1381210,
    "ts": 1514658025961.324, "dur": 6.903,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026043.344, "dur": 2.053,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026060.725, "dur": 1.432,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026074.098, "dur": 2.440,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026088.440, "dur": 1.556,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026099.883, "dur": 1.146,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026112.520, "dur": 1.299,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026125.317, "dur": 2.737,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026139.266, "dur": 1.379,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026151.676, "dur": 1.505,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026164.390, "dur": 1.207,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026175.513, "dur": 3.200,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026189.945, "dur": 1.191,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026201.900, "dur": 1.221,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026212.396, "dur": 1.193,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026223.870, "dur": 2.692,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026236.872, "dur": 1.487,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026248.543, "dur": 1.638,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026259.688, "dur": 1.241,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026270.465, "dur": 2.451,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026283.252, "dur": 0.977,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026294.161, "dur": 1.094,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026304.612, "dur": 1.115,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026315.552, "dur": 2.378,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026327.563, "dur": 1.332,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026338.210, "dur": 1.695,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026350.208, "dur": 1.832,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026361.546, "dur": 2.104,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026373.196, "dur": 1.325,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026384.706, "dur": 1.995,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026395.438, "dur": 1.359,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026406.470, "dur": 2.605,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026419.285, "dur": 1.281,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026430.580, "dur": 1.542,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026442.598, "dur": 1.591,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026454.111, "dur": 2.078,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026467.252, "dur": 1.228,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026478.406, "dur": 1.233,
    "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": 1336753, "tid": 1381210,
    "ts": 1514658026489.711, "dur": 1.424,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655937706.304, "dur": 2107848.336,
    "args": {
      "External id": 2478081,"Record function id": 0, "Ev Idx": 5784
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "Optimizer.zero_grad#AdamW.zero_grad", "pid": 1336753, "tid": 1336753,
    "ts": 1514655937742.809, "dur": 590.344,
    "args": {
      "External id": 2478082,"Record function id": 0, "Ev Idx": 5785
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "enumerate(DataLoader)#_StatefulMultiProcessingDataLoaderIter.__next__", "pid": 1336753, "tid": 1336753,
    "ts": 1514655938379.608, "dur": 2071.542,
    "args": {
      "External id": 2478083,"Record function id": 0, "Ev Idx": 5786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514655939454.023, "dur": 7.882,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655939508.357, "dur": 7.951,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655939944.533, "dur": 2.305,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655939954.331, "dur": 2.537,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655940340.378, "dur": 1.929,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655940349.229, "dur": 1.875,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655941071.709, "dur": 14.091,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655941079.531, "dur": 2.325,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655941087.352, "dur": 4.351,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655941089.396, "dur": 1.268,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655941118.481, "dur": 576.919,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655941125.284, "dur": 569.082,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655941138.320, "dur": 9.986,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655941150.980, "dur": 540.772,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655941160.655, "dur": 0.400,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655941163.670, "dur": 8.434,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655941168.621, "dur": 3.323,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655941171.050, "dur": 0.626,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655941173.892, "dur": 224.842,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655941176.375, "dur": 221.839,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655941178.556, "dur": 18.080,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655941180.658, "dur": 15.439,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655941197.414, "dur": 200.364,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655941400.700, "dur": 278.942,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655941726.502, "dur": 652.207,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655941729.789, "dur": 647.953,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655941742.817, "dur": 16.439,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655941761.618, "dur": 610.344,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655942412.244, "dur": 62.283,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655942419.059, "dur": 6.472,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655942428.338, "dur": 45.700,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655942433.896, "dur": 8.680,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655942490.262, "dur": 118.519,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655942495.754, "dur": 8.788,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655942501.873, "dur": 2.275,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655942506.017, "dur": 6.316,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655942515.614, "dur": 3.997,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655942525.108, "dur": 26.292,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655942549.659, "dur": 1.084,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655942552.835, "dur": 7.796,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655942559.104, "dur": 1.356,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655942562.922, "dur": 5.631,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655942565.025, "dur": 3.333,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655942567.051, "dur": 1.094,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655942570.182, "dur": 37.612,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655942620.832, "dur": 41.011,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655942622.725, "dur": 38.872,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655942632.115, "dur": 4.871,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655942638.093, "dur": 22.868,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655942789.745, "dur": 272.826,
    "args": {
      "External id": 2478133,"Record function id": 0, "Ev Idx": 5836
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::inputs_to_device", "pid": 1336753, "tid": 1336753,
    "ts": 1514655942909.088, "dur": 134.797,
    "args": {
      "External id": 2478134,"Record function id": 0, "Ev Idx": 5837
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336753, "tid": 1336753,
    "ts": 1514655943075.779, "dur": 60.708,
    "args": {
      "External id": 2478135,"Record function id": 0, "Ev Idx": 5838
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward", "pid": 1336753, "tid": 1336753,
    "ts": 1514655943148.682, "dur": 9398.434,
    "args": {
      "External id": 2478136,"Record function id": 0, "Ev Idx": 5839
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather", "pid": 1336753, "tid": 1336753,
    "ts": 1514655943160.357, "dur": 1924.808,
    "args": {
      "External id": 2478137,"Record function id": 0, "Ev Idx": 5840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514655943306.673, "dur": 13.415,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943343.635, "dur": 206.707,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943350.780, "dur": 2.837,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943361.481, "dur": 0.915,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943363.831, "dur": 0.452,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943365.432, "dur": 2.582,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943374.187, "dur": 0.504,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943376.238, "dur": 0.652,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943378.169, "dur": 3.034,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943384.920, "dur": 0.644,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943386.814, "dur": 0.653,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943393.610, "dur": 0.722,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943395.725, "dur": 0.648,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943397.359, "dur": 2.398,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943405.289, "dur": 0.507,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943406.949, "dur": 0.502,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943408.614, "dur": 2.720,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943415.182, "dur": 0.876,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943417.594, "dur": 0.468,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943424.618, "dur": 0.850,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943426.600, "dur": 0.763,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943428.668, "dur": 2.280,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943437.108, "dur": 0.539,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943438.674, "dur": 0.711,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943440.521, "dur": 2.777,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943447.028, "dur": 0.894,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943449.151, "dur": 0.885,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943456.000, "dur": 0.695,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943457.736, "dur": 0.671,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943459.542, "dur": 2.224,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943468.246, "dur": 0.480,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943473.076, "dur": 0.488,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943474.804, "dur": 3.103,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943481.386, "dur": 0.677,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943483.225, "dur": 0.691,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943490.333, "dur": 0.713,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943492.352, "dur": 0.626,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943494.120, "dur": 1.986,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943502.551, "dur": 0.671,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943504.358, "dur": 0.420,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943505.937, "dur": 2.684,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943589.241, "dur": 131.664,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943846.027, "dur": 448.840,
    "args": {
      "External id": 2478180,"Record function id": 0, "Concrete Inputs": ["", "", "42076416", "8", "0", "15", ""], "Input type": ["TensorList", "", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], []], "Ev Idx": 5883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514655943871.860, "dur": 12.603,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943893.198, "dur": 20.459,
    "args": {
      "External id": 2478182,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514655943898.882, "dur": 14.209,
    "args": {
      "External id": 2478183,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "42076416", "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": 1336753, "tid": 1336753,
    "ts": 1514655943907.270, "dur": 1.545,
    "args": {
      "External id": 2478184,"Record function id": 0, "Concrete Inputs": ["", "[42076416]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514655943927.639, "dur": 211.852,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655943930.820, "dur": 1.053,
    "args": {
      "External id": 2478186,"Record function id": 0, "Concrete Inputs": ["", "[8192000]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514655943934.683, "dur": 0.812,
    "args": {
      "External id": 2478187,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "8192000"], "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": 1336753, "tid": 1336753,
    "ts": 1514655943939.498, "dur": 2.780,
    "args": {
      "External id": 2478188,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "8192256"], "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": 1336753, "tid": 1336753,
    "ts": 1514655943943.589, "dur": 1.693,
    "args": {
      "External id": 2478189,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "8716544"], "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": 1336753, "tid": 1336753,
    "ts": 1514655943948.866, "dur": 0.516,
    "args": {
      "External id": 2478190,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "9240832"], "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": 1336753, "tid": 1336753,
    "ts": 1514655943950.498, "dur": 0.626,
    "args": {
      "External id": 2478191,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "9765120"], "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": 1336753, "tid": 1336753,
    "ts": 1514655943953.921, "dur": 0.599,
    "args": {
      "External id": 2478192,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "10289408"], "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": 1336753, "tid": 1336753,
    "ts": 1514655943957.952, "dur": 0.881,
    "args": {
      "External id": 2478193,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "10289664"], "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": 1336753, "tid": 1336753,
    "ts": 1514655943959.966, "dur": 1.837,
    "args": {
      "External id": 2478194,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "11731456"], "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": 1336753, "tid": 1336753,
    "ts": 1514655943962.977, "dur": 0.442,
    "args": {
      "External id": 2478195,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "13173248"], "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": 1336753, "tid": 1336753,
    "ts": 1514655943968.113, "dur": 2.582,
    "args": {
      "External id": 2478196,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "14615040"], "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": 1336753, "tid": 1336753,
    "ts": 1514655943972.959, "dur": 0.771,
    "args": {
      "External id": 2478197,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "14615296"], "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": 1336753, "tid": 1336753,
    "ts": 1514655943977.131, "dur": 2.640,
    "args": {
      "External id": 2478198,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "15139584"], "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": 1336753, "tid": 1336753,
    "ts": 1514655943981.077, "dur": 0.310,
    "args": {
      "External id": 2478199,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "15663872"], "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": 1336753, "tid": 1336753,
    "ts": 1514655943982.610, "dur": 0.679,
    "args": {
      "External id": 2478200,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "16188160"], "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": 1336753, "tid": 1336753,
    "ts": 1514655944032.077, "dur": 1.093,
    "args": {
      "External id": 2478201,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "16712448"], "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": 1336753, "tid": 1336753,
    "ts": 1514655944035.134, "dur": 0.810,
    "args": {
      "External id": 2478202,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "16712704"], "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": 1336753, "tid": 1336753,
    "ts": 1514655944037.143, "dur": 0.626,
    "args": {
      "External id": 2478203,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "18154496"], "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": 1336753, "tid": 1336753,
    "ts": 1514655944044.321, "dur": 3.137,
    "args": {
      "External id": 2478204,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "19596288"], "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": 1336753, "tid": 1336753,
    "ts": 1514655944048.610, "dur": 0.481,
    "args": {
      "External id": 2478205,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "21038080"], "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": 1336753, "tid": 1336753,
    "ts": 1514655944053.030, "dur": 2.464,
    "args": {
      "External id": 2478206,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "21038336"], "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": 1336753, "tid": 1336753,
    "ts": 1514655944056.687, "dur": 0.887,
    "args": {
      "External id": 2478207,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "21562624"], "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": 1336753, "tid": 1336753,
    "ts": 1514655944058.791, "dur": 0.826,
    "args": {
      "External id": 2478208,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "22086912"], "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": 1336753, "tid": 1336753,
    "ts": 1514655944065.183, "dur": 0.508,
    "args": {
      "External id": 2478209,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "22611200"], "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": 1336753, "tid": 1336753,
    "ts": 1514655944066.943, "dur": 0.447,
    "args": {
      "External id": 2478210,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "23135488"], "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": 1336753, "tid": 1336753,
    "ts": 1514655944068.790, "dur": 0.440,
    "args": {
      "External id": 2478211,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "23135744"], "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": 1336753, "tid": 1336753,
    "ts": 1514655944075.428, "dur": 1.628,
    "args": {
      "External id": 2478212,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "24577536"], "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": 1336753, "tid": 1336753,
    "ts": 1514655944078.172, "dur": 0.834,
    "args": {
      "External id": 2478213,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "26019328"], "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": 1336753, "tid": 1336753,
    "ts": 1514655944082.620, "dur": 3.137,
    "args": {
      "External id": 2478214,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "27461120"], "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": 1336753, "tid": 1336753,
    "ts": 1514655944086.971, "dur": 0.625,
    "args": {
      "External id": 2478215,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "27461376"], "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": 1336753, "tid": 1336753,
    "ts": 1514655944088.891, "dur": 0.666,
    "args": {
      "External id": 2478216,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "27985664"], "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": 1336753, "tid": 1336753,
    "ts": 1514655944095.538, "dur": 0.523,
    "args": {
      "External id": 2478217,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "28509952"], "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": 1336753, "tid": 1336753,
    "ts": 1514655944097.145, "dur": 0.451,
    "args": {
      "External id": 2478218,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "29034240"], "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": 1336753, "tid": 1336753,
    "ts": 1514655944098.733, "dur": 0.500,
    "args": {
      "External id": 2478219,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "29558528"], "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": 1336753, "tid": 1336753,
    "ts": 1514655944105.421, "dur": 1.289,
    "args": {
      "External id": 2478220,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "29558784"], "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": 1336753, "tid": 1336753,
    "ts": 1514655944107.753, "dur": 0.671,
    "args": {
      "External id": 2478221,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "31000576"], "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": 1336753, "tid": 1336753,
    "ts": 1514655944111.764, "dur": 2.467,
    "args": {
      "External id": 2478222,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "32442368"], "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": 1336753, "tid": 1336753,
    "ts": 1514655944115.357, "dur": 0.454,
    "args": {
      "External id": 2478223,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "33884160"], "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": 1336753, "tid": 1336753,
    "ts": 1514655944117.118, "dur": 0.489,
    "args": {
      "External id": 2478224,"Record function id": 0, "Concrete Inputs": ["", "[8192000]", "[1]", "33884416"], "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": 1336753, "tid": 1336753,
    "ts": 1514655944175.893, "dur": 94.967,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655944382.302, "dur": 501.257,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655944426.506, "dur": 448.674,
    "args": {
      "External id": 2478227,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1336753,
    "ts": 1514655944443.036, "dur": 420.697,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655944921.266, "dur": 3.375,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out", "pid": 1336753, "tid": 1336753,
    "ts": 1514655945107.712, "dur": 7155.249,
    "args": {
      "External id": 2478230,"Record function id": 0, "Ev Idx": 5933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514655945376.887, "dur": 8.954,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945391.552, "dur": 1.604,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945396.371, "dur": 1.220,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945400.957, "dur": 3.303,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945407.161, "dur": 1.352,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945414.104, "dur": 1.229,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945417.471, "dur": 1.532,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945421.239, "dur": 3.211,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945429.443, "dur": 1.267,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945435.196, "dur": 1.153,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945438.880, "dur": 0.895,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945442.043, "dur": 2.557,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945447.007, "dur": 1.338,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945452.824, "dur": 0.985,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945456.079, "dur": 1.260,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945459.852, "dur": 2.450,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945467.125, "dur": 1.304,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945473.510, "dur": 1.399,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945477.437, "dur": 0.944,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945480.681, "dur": 2.391,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945485.440, "dur": 1.360,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945491.991, "dur": 1.346,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945495.615, "dur": 1.314,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945499.275, "dur": 2.390,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945549.189, "dur": 2.603,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945559.077, "dur": 1.198,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945562.723, "dur": 1.143,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945566.596, "dur": 2.841,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945571.877, "dur": 1.234,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945577.809, "dur": 1.129,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945581.576, "dur": 1.301,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945585.381, "dur": 2.419,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945592.431, "dur": 1.631,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945598.997, "dur": 1.188,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945602.999, "dur": 0.885,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945606.377, "dur": 2.022,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945610.907, "dur": 1.014,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945616.675, "dur": 1.022,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945619.777, "dur": 1.250,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945623.443, "dur": 2.224,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945673.349, "dur": 6499.556,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945708.944, "dur": 6446.783,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945746.436, "dur": 7.731,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945763.533, "dur": 6316.494,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945766.417, "dur": 6312.264,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945776.086, "dur": 9.876,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655945788.186, "dur": 6283.925,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655952667.790, "dur": 49.284,
    "args": {
      "External id": 2478278,"Record function id": 0, "Ev Idx": 5981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/0", "pid": 1336753, "tid": 1336753,
    "ts": 1514655952720.132, "dur": 389.628,
    "args": {
      "External id": 2478279,"Record function id": 0, "Ev Idx": 5982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514655952779.467, "dur": 311.450,
    "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": 1336753, "tid": 1336753, "ts": 1514655952779.467,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_embedding_0", "pid": 1336753, "tid": 1336753,
    "ts": 1514655952880.306, "dur": 82.838,
    "args": {
      "External id": 2478281,"kernel_hash": "cepuvqida76mp3inhfyv5xcp3dgr5kc6gyyaupvhufqptth43nyf", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/ep/cepuvqida76mp3inhfyv5xcp3dgr5kc6gyyaupvhufqptth43nyf.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": 1336753, "tid": 1336753,
    "ts": 1514655953207.655, "dur": 88.610,
    "args": {
      "External id": 2478282,"Record function id": 0, "Ev Idx": 5985
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.0)", "pid": 1336753, "tid": 1336753,
    "ts": 1514655953314.676, "dur": 7636.858,
    "args": {
      "External id": 2478283,"Record function id": 0, "Ev Idx": 5986
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.0)", "pid": 1336753, "tid": 1336753,
    "ts": 1514655953328.992, "dur": 1233.068,
    "args": {
      "External id": 2478284,"Record function id": 0, "Ev Idx": 5987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514655953435.386, "dur": 17.304,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655953472.758, "dur": 79.513,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655953485.161, "dur": 2.993,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655953490.963, "dur": 0.762,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655953493.214, "dur": 2.820,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655953499.914, "dur": 0.475,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655953501.802, "dur": 0.274,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655953509.275, "dur": 0.458,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655953510.995, "dur": 0.539,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655953512.866, "dur": 2.577,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655953521.743, "dur": 0.306,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655953567.997, "dur": 64.771,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655953688.304, "dur": 198.667,
    "args": {
      "External id": 2478297,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514655953706.375, "dur": 10.838,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655953725.218, "dur": 17.823,
    "args": {
      "External id": 2478299,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514655953730.612, "dur": 11.887,
    "args": {
      "External id": 2478300,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1336753,
    "ts": 1514655953738.908, "dur": 1.328,
    "args": {
      "External id": 2478301,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514655953758.372, "dur": 44.760,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655953761.430, "dur": 1.110,
    "args": {
      "External id": 2478303,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514655953767.156, "dur": 0.760,
    "args": {
      "External id": 2478304,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1336753,
    "ts": 1514655953772.047, "dur": 0.271,
    "args": {
      "External id": 2478305,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1336753,
    "ts": 1514655953776.490, "dur": 0.474,
    "args": {
      "External id": 2478306,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1336753,
    "ts": 1514655953778.438, "dur": 4.707,
    "args": {
      "External id": 2478307,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1336753,
    "ts": 1514655953784.842, "dur": 0.322,
    "args": {
      "External id": 2478308,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1336753,
    "ts": 1514655953789.586, "dur": 0.286,
    "args": {
      "External id": 2478309,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1336753,
    "ts": 1514655953794.093, "dur": 0.308,
    "args": {
      "External id": 2478310,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1336753,
    "ts": 1514655953795.806, "dur": 0.735,
    "args": {
      "External id": 2478311,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1336753,
    "ts": 1514655953825.685, "dur": 45.535,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655953973.073, "dur": 423.363,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655954053.975, "dur": 334.095,
    "args": {
      "External id": 2478314,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1336753,
    "ts": 1514655954070.195, "dur": 308.888,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655954430.102, "dur": 4.071,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.0)", "pid": 1336753, "tid": 1336753,
    "ts": 1514655954601.479, "dur": 6040.991,
    "args": {
      "External id": 2478317,"Record function id": 0, "Ev Idx": 6020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514655954751.313, "dur": 8.992,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655954766.208, "dur": 1.634,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655954770.752, "dur": 1.438,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655954777.425, "dur": 2.473,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655954782.332, "dur": 1.440,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655954785.726, "dur": 1.436,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655954789.850, "dur": 1.329,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655954798.798, "dur": 2.726,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655954803.861, "dur": 1.282,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655954807.263, "dur": 1.385,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655954832.785, "dur": 5735.353,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655954857.407, "dur": 5665.655,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655954893.400, "dur": 18.236,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655954918.733, "dur": 5541.186,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655954922.246, "dur": 5536.455,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655954929.604, "dur": 10.405,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655954942.686, "dur": 5509.187,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655960861.362, "dur": 48.302,
    "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": 1336753, "tid": 1336753, "ts": 1514655960861.362,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336753, "tid": 1336753,
    "ts": 1514655960884.865, "dur": 17.653,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655960893.810, "dur": 8.241,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655961077.178, "dur": 115.917,
    "args": {
      "External id": 2478338,"Record function id": 0, "Ev Idx": 6041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336753, "tid": 1336753,
    "ts": 1514655961197.307, "dur": 2000.653,
    "args": {
      "External id": 2478339,"Record function id": 0, "Ev Idx": 6042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514655961262.707, "dur": 1913.201,
    "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": 1336753, "tid": 1336753, "ts": 1514655961262.707,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336753, "tid": 1336753,
    "ts": 1514655961365.853, "dur": 73.254,
    "args": {
      "External id": 2478341,"kernel_hash": "cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/ml/cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc.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": 1336753, "tid": 1336753,
    "ts": 1514655961461.363, "dur": 164.497,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655961654.457, "dur": 70.617,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655961742.159, "dur": 54.948,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655961848.477, "dur": 48.107,
    "args": {
      "External id": 2478345,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514655961931.143, "dur": 30.020,
    "args": {
      "External id": 2478346,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514655962052.983, "dur": 209.965,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655962127.055, "dur": 21.379,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655962134.285, "dur": 12.356,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655962152.170, "dur": 5.452,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655962162.964, "dur": 1.793,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655962169.441, "dur": 5.746,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655962283.161, "dur": 112.979,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655962461.646, "dur": 51.208,
    "args": {
      "External id": 2478354,"kernel_hash": "cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/w7/cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5.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": 1336753, "tid": 1336753,
    "ts": 1514655962563.287, "dur": 93.623,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655962687.262, "dur": 62.204,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655962787.532, "dur": 41.664,
    "args": {
      "External id": 2478357,"kernel_hash": "czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/zh/czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha.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": 1336753, "tid": 1336753,
    "ts": 1514655962840.891, "dur": 64.299,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655962953.913, "dur": 99.439,
    "args": {
      "External id": 2478359,"kernel_hash": "cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/nz/cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4.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": 1336753, "tid": 1336753,
    "ts": 1514655963284.598, "dur": 115.070,
    "args": {
      "External id": 2478360,"Record function id": 0, "Ev Idx": 6063
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336753, "tid": 1336753,
    "ts": 1514655963508.696, "dur": 110.018,
    "args": {
      "External id": 2478361,"Record function id": 0, "Ev Idx": 6064
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.1)", "pid": 1336753, "tid": 1336753,
    "ts": 1514655963636.942, "dur": 17139.947,
    "args": {
      "External id": 2478362,"Record function id": 0, "Ev Idx": 6065
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.1)", "pid": 1336753, "tid": 1336753,
    "ts": 1514655963652.012, "dur": 1448.563,
    "args": {
      "External id": 2478363,"Record function id": 0, "Ev Idx": 6066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514655963759.307, "dur": 15.715,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655963796.961, "dur": 60.871,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655963806.622, "dur": 5.337,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655963814.574, "dur": 0.804,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655963819.321, "dur": 0.880,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655963824.365, "dur": 0.561,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655963826.760, "dur": 2.746,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655963834.008, "dur": 0.520,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655963838.850, "dur": 0.670,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655963841.320, "dur": 0.557,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655963845.780, "dur": 3.169,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655963872.060, "dur": 67.369,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655964056.192, "dur": 215.515,
    "args": {
      "External id": 2478376,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514655964079.008, "dur": 11.426,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655964099.047, "dur": 15.995,
    "args": {
      "External id": 2478378,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514655964104.327, "dur": 10.098,
    "args": {
      "External id": 2478379,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1336753,
    "ts": 1514655964109.620, "dur": 1.476,
    "args": {
      "External id": 2478380,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514655964128.355, "dur": 49.517,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655964134.920, "dur": 1.042,
    "args": {
      "External id": 2478382,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514655964137.868, "dur": 2.820,
    "args": {
      "External id": 2478383,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1336753,
    "ts": 1514655964142.099, "dur": 3.294,
    "args": {
      "External id": 2478384,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1336753,
    "ts": 1514655964149.479, "dur": 0.384,
    "args": {
      "External id": 2478385,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1336753,
    "ts": 1514655964151.436, "dur": 0.905,
    "args": {
      "External id": 2478386,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1336753,
    "ts": 1514655964156.026, "dur": 0.559,
    "args": {
      "External id": 2478387,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1336753,
    "ts": 1514655964161.478, "dur": 0.276,
    "args": {
      "External id": 2478388,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1336753,
    "ts": 1514655964163.394, "dur": 0.724,
    "args": {
      "External id": 2478389,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1336753,
    "ts": 1514655964170.824, "dur": 0.349,
    "args": {
      "External id": 2478390,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1336753,
    "ts": 1514655964200.605, "dur": 53.756,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655964357.392, "dur": 529.575,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655964401.304, "dur": 476.234,
    "args": {
      "External id": 2478393,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1336753,
    "ts": 1514655964415.217, "dur": 452.346,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655964926.308, "dur": 4.310,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.1)", "pid": 1336753, "tid": 1336753,
    "ts": 1514655965135.296, "dur": 15417.016,
    "args": {
      "External id": 2478396,"Record function id": 0, "Ev Idx": 6099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514655965276.575, "dur": 9.127,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655965291.448, "dur": 1.758,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655965296.515, "dur": 1.551,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655965301.022, "dur": 1.302,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655965304.762, "dur": 1.240,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655965311.009, "dur": 1.699,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655965318.129, "dur": 1.360,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655965321.707, "dur": 3.955,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655965328.050, "dur": 1.868,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655965334.373, "dur": 1.124,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655965359.864, "dur": 15111.978,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655965383.646, "dur": 15075.251,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655965425.535, "dur": 18.740,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655965449.279, "dur": 14959.607,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655965453.038, "dur": 14954.484,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655965459.965, "dur": 7.748,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655965470.290, "dur": 14930.861,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655980712.998, "dur": 36.950,
    "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": 1336753, "tid": 1336753, "ts": 1514655980712.998,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336753, "tid": 1336753,
    "ts": 1514655980730.742, "dur": 14.355,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655980737.797, "dur": 6.879,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655980824.154, "dur": 82.489,
    "args": {
      "External id": 2478417,"Record function id": 0, "Ev Idx": 6120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336753, "tid": 1336753,
    "ts": 1514655980908.800, "dur": 1166.415,
    "args": {
      "External id": 2478418,"Record function id": 0, "Ev Idx": 6121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514655980950.115, "dur": 1110.145,
    "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": 1336753, "tid": 1336753, "ts": 1514655980950.115,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336753, "tid": 1336753,
    "ts": 1514655981076.655, "dur": 50.717,
    "args": {
      "External id": 2478420,"kernel_hash": "cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/ml/cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc.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": 1336753, "tid": 1336753,
    "ts": 1514655981143.814, "dur": 102.399,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655981254.073, "dur": 37.727,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655981297.318, "dur": 30.338,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655981361.593, "dur": 34.478,
    "args": {
      "External id": 2478424,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514655981412.428, "dur": 19.223,
    "args": {
      "External id": 2478425,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514655981448.715, "dur": 159.167,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655981500.530, "dur": 12.936,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655981505.632, "dur": 7.004,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655981519.382, "dur": 4.764,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655981525.458, "dur": 20.954,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655981550.111, "dur": 3.186,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655981619.093, "dur": 50.308,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655981700.924, "dur": 27.960,
    "args": {
      "External id": 2478433,"kernel_hash": "cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/w7/cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5.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": 1336753, "tid": 1336753,
    "ts": 1514655981738.624, "dur": 39.697,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655981784.218, "dur": 34.275,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655981843.826, "dur": 23.959,
    "args": {
      "External id": 2478436,"kernel_hash": "czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/zh/czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha.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": 1336753, "tid": 1336753,
    "ts": 1514655981873.693, "dur": 33.589,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655981924.543, "dur": 22.254,
    "args": {
      "External id": 2478438,"kernel_hash": "cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/nz/cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4.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": 1336753, "tid": 1336753,
    "ts": 1514655982144.153, "dur": 92.250,
    "args": {
      "External id": 2478439,"Record function id": 0, "Ev Idx": 6142
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336753, "tid": 1336753,
    "ts": 1514655982341.745, "dur": 62.405,
    "args": {
      "External id": 2478440,"Record function id": 0, "Ev Idx": 6143
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.2)", "pid": 1336753, "tid": 1336753,
    "ts": 1514655982422.574, "dur": 18406.908,
    "args": {
      "External id": 2478441,"Record function id": 0, "Ev Idx": 6144
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.2)", "pid": 1336753, "tid": 1336753,
    "ts": 1514655982435.948, "dur": 1265.505,
    "args": {
      "External id": 2478442,"Record function id": 0, "Ev Idx": 6145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514655982548.176, "dur": 12.586,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655982583.159, "dur": 55.590,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655982589.740, "dur": 2.776,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655982597.713, "dur": 0.547,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655982603.900, "dur": 0.759,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655982606.232, "dur": 0.751,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655982611.056, "dur": 0.630,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655982615.947, "dur": 0.822,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655982618.232, "dur": 4.596,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655982626.740, "dur": 0.695,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655982628.947, "dur": 0.476,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655982651.444, "dur": 60.619,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655982762.542, "dur": 159.515,
    "args": {
      "External id": 2478455,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514655982780.344, "dur": 6.889,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655982794.285, "dur": 12.010,
    "args": {
      "External id": 2478457,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514655982799.030, "dur": 6.667,
    "args": {
      "External id": 2478458,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1336753,
    "ts": 1514655982802.989, "dur": 0.781,
    "args": {
      "External id": 2478459,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514655982816.249, "dur": 45.305,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655982821.095, "dur": 2.836,
    "args": {
      "External id": 2478461,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514655982825.488, "dur": 0.607,
    "args": {
      "External id": 2478462,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1336753,
    "ts": 1514655982827.282, "dur": 0.507,
    "args": {
      "External id": 2478463,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1336753,
    "ts": 1514655982834.130, "dur": 1.725,
    "args": {
      "External id": 2478464,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1336753,
    "ts": 1514655982837.461, "dur": 0.557,
    "args": {
      "External id": 2478465,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1336753,
    "ts": 1514655982841.873, "dur": 0.491,
    "args": {
      "External id": 2478466,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1336753,
    "ts": 1514655982846.195, "dur": 0.286,
    "args": {
      "External id": 2478467,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1336753,
    "ts": 1514655982847.818, "dur": 0.356,
    "args": {
      "External id": 2478468,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1336753,
    "ts": 1514655982852.636, "dur": 2.610,
    "args": {
      "External id": 2478469,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1336753,
    "ts": 1514655982875.165, "dur": 35.449,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655983040.738, "dur": 519.086,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655983081.955, "dur": 470.063,
    "args": {
      "External id": 2478472,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1336753,
    "ts": 1514655983098.122, "dur": 424.693,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655983597.558, "dur": 4.046,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.2)", "pid": 1336753, "tid": 1336753,
    "ts": 1514655983729.391, "dur": 16759.409,
    "args": {
      "External id": 2478475,"Record function id": 0, "Ev Idx": 6178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514655983857.918, "dur": 7.854,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655983870.559, "dur": 1.718,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655983874.704, "dur": 2.804,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655983880.101, "dur": 1.441,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655983883.835, "dur": 1.231,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655983887.308, "dur": 1.222,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655983893.355, "dur": 1.411,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655983897.982, "dur": 2.757,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655983903.155, "dur": 1.505,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655983906.720, "dur": 1.142,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655983935.114, "dur": 16477.331,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655983954.507, "dur": 16443.605,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655984025.796, "dur": 19.275,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655984050.122, "dur": 16285.092,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655984053.566, "dur": 16280.464,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655984060.758, "dur": 7.929,
    "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": 1336753, "tid": 1336753,
    "ts": 1514655984071.062, "dur": 16255.810,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656000736.947, "dur": 47.173,
    "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": 1336753, "tid": 1336753, "ts": 1514656000736.947,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336753, "tid": 1336753,
    "ts": 1514656000761.864, "dur": 15.417,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656000767.802, "dur": 8.728,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656000888.400, "dur": 180.010,
    "args": {
      "External id": 2478496,"Record function id": 0, "Ev Idx": 6199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336753, "tid": 1336753,
    "ts": 1514656001073.497, "dur": 1823.394,
    "args": {
      "External id": 2478497,"Record function id": 0, "Ev Idx": 6200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656001138.516, "dur": 1736.729,
    "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": 1336753, "tid": 1336753, "ts": 1514656001138.516,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336753, "tid": 1336753,
    "ts": 1514656001241.262, "dur": 67.435,
    "args": {
      "External id": 2478499,"kernel_hash": "cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/ml/cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc.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": 1336753, "tid": 1336753,
    "ts": 1514656001330.439, "dur": 121.765,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656001469.681, "dur": 91.410,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656001583.729, "dur": 63.717,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656001698.804, "dur": 42.070,
    "args": {
      "External id": 2478503,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656001773.584, "dur": 29.274,
    "args": {
      "External id": 2478504,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656001833.649, "dur": 257.701,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656001897.763, "dur": 19.340,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656001907.427, "dur": 8.324,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656001920.308, "dur": 7.242,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656001929.332, "dur": 1.812,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656001934.445, "dur": 5.893,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656002114.986, "dur": 110.328,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656002285.377, "dur": 47.628,
    "args": {
      "External id": 2478512,"kernel_hash": "cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/w7/cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5.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": 1336753, "tid": 1336753,
    "ts": 1514656002348.384, "dur": 73.990,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656002438.069, "dur": 60.746,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656002562.256, "dur": 47.372,
    "args": {
      "External id": 2478515,"kernel_hash": "czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/zh/czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha.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": 1336753, "tid": 1336753,
    "ts": 1514656002623.370, "dur": 69.774,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656002736.039, "dur": 31.935,
    "args": {
      "External id": 2478517,"kernel_hash": "cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/nz/cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4.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": 1336753, "tid": 1336753,
    "ts": 1514656003046.661, "dur": 113.194,
    "args": {
      "External id": 2478518,"Record function id": 0, "Ev Idx": 6221
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336753, "tid": 1336753,
    "ts": 1514656003272.746, "dur": 72.040,
    "args": {
      "External id": 2478519,"Record function id": 0, "Ev Idx": 6222
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.3)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656003362.346, "dur": 17084.056,
    "args": {
      "External id": 2478520,"Record function id": 0, "Ev Idx": 6223
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.3)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656003381.144, "dur": 1464.768,
    "args": {
      "External id": 2478521,"Record function id": 0, "Ev Idx": 6224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514656003488.642, "dur": 14.612,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656003524.124, "dur": 88.868,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656003559.018, "dur": 3.098,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656003570.599, "dur": 0.755,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656003572.816, "dur": 0.717,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656003580.317, "dur": 0.620,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656003587.875, "dur": 0.602,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656003589.802, "dur": 0.635,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656003595.457, "dur": 4.013,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656003601.101, "dur": 0.441,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656003602.754, "dur": 0.354,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656003632.984, "dur": 68.888,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656003758.998, "dur": 308.220,
    "args": {
      "External id": 2478534,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656003781.454, "dur": 10.397,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656003800.152, "dur": 14.185,
    "args": {
      "External id": 2478536,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656003805.172, "dur": 8.579,
    "args": {
      "External id": 2478537,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1336753,
    "ts": 1514656003810.584, "dur": 1.101,
    "args": {
      "External id": 2478538,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656003829.814, "dur": 47.435,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656003832.539, "dur": 3.544,
    "args": {
      "External id": 2478540,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656003837.758, "dur": 0.418,
    "args": {
      "External id": 2478541,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1336753,
    "ts": 1514656003843.078, "dur": 1.106,
    "args": {
      "External id": 2478542,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1336753,
    "ts": 1514656003848.869, "dur": 2.025,
    "args": {
      "External id": 2478543,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1336753,
    "ts": 1514656003854.704, "dur": 0.637,
    "args": {
      "External id": 2478544,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1336753,
    "ts": 1514656003856.616, "dur": 0.401,
    "args": {
      "External id": 2478545,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1336753,
    "ts": 1514656003860.394, "dur": 0.556,
    "args": {
      "External id": 2478546,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1336753,
    "ts": 1514656003864.800, "dur": 0.504,
    "args": {
      "External id": 2478547,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1336753,
    "ts": 1514656003866.630, "dur": 2.848,
    "args": {
      "External id": 2478548,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1336753,
    "ts": 1514656003944.072, "dur": 103.175,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656004158.383, "dur": 527.505,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656004203.633, "dur": 473.154,
    "args": {
      "External id": 2478551,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1336753,
    "ts": 1514656004225.732, "dur": 441.147,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656004731.451, "dur": 4.246,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.3)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656004877.422, "dur": 15337.905,
    "args": {
      "External id": 2478554,"Record function id": 0, "Ev Idx": 6257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656005067.052, "dur": 9.552,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656005082.853, "dur": 1.748,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656005087.516, "dur": 3.548,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656005094.520, "dur": 1.417,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656005098.056, "dur": 1.543,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656005102.072, "dur": 1.263,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656005108.542, "dur": 1.186,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656005112.536, "dur": 3.766,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656005118.783, "dur": 1.247,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656005122.360, "dur": 1.328,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656005157.203, "dur": 14990.680,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656005180.255, "dur": 14954.055,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656005212.835, "dur": 18.882,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656005235.622, "dur": 14841.935,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656005239.167, "dur": 14837.034,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656005249.543, "dur": 9.062,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656005261.160, "dur": 14807.755,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656020370.600, "dur": 47.347,
    "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": 1336753, "tid": 1336753, "ts": 1514656020370.600,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336753, "tid": 1336753,
    "ts": 1514656020400.966, "dur": 12.249,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656020406.352, "dur": 6.640,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656020491.408, "dur": 93.782,
    "args": {
      "External id": 2478575,"Record function id": 0, "Ev Idx": 6278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336753, "tid": 1336753,
    "ts": 1514656020591.343, "dur": 1220.651,
    "args": {
      "External id": 2478576,"Record function id": 0, "Ev Idx": 6279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656020634.907, "dur": 1162.725,
    "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": 1336753, "tid": 1336753, "ts": 1514656020634.907,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336753, "tid": 1336753,
    "ts": 1514656020706.304, "dur": 49.608,
    "args": {
      "External id": 2478578,"kernel_hash": "cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/ml/cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc.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": 1336753, "tid": 1336753,
    "ts": 1514656020767.448, "dur": 105.532,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656020887.700, "dur": 39.780,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656020933.446, "dur": 31.169,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656021032.036, "dur": 32.294,
    "args": {
      "External id": 2478582,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656021083.774, "dur": 15.716,
    "args": {
      "External id": 2478583,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656021119.404, "dur": 132.974,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656021171.485, "dur": 12.238,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656021176.508, "dur": 6.319,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656021186.224, "dur": 5.203,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656021192.912, "dur": 1.161,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656021199.275, "dur": 3.797,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656021263.356, "dur": 54.837,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656021351.592, "dur": 28.692,
    "args": {
      "External id": 2478591,"kernel_hash": "cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/w7/cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5.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": 1336753, "tid": 1336753,
    "ts": 1514656021389.237, "dur": 40.779,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656021513.789, "dur": 58.395,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656021607.338, "dur": 30.969,
    "args": {
      "External id": 2478594,"kernel_hash": "czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/zh/czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha.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": 1336753, "tid": 1336753,
    "ts": 1514656021644.905, "dur": 38.664,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656021701.555, "dur": 21.969,
    "args": {
      "External id": 2478596,"kernel_hash": "cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/nz/cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4.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": 1336753, "tid": 1336753,
    "ts": 1514656021875.900, "dur": 74.600,
    "args": {
      "External id": 2478597,"Record function id": 0, "Ev Idx": 6300
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336753, "tid": 1336753,
    "ts": 1514656022067.523, "dur": 49.858,
    "args": {
      "External id": 2478598,"Record function id": 0, "Ev Idx": 6301
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.4)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656022127.122, "dur": 17939.759,
    "args": {
      "External id": 2478599,"Record function id": 0, "Ev Idx": 6302
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.4)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656022138.138, "dur": 1070.827,
    "args": {
      "External id": 2478600,"Record function id": 0, "Ev Idx": 6303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514656022217.583, "dur": 9.015,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656022240.924, "dur": 44.300,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656022248.453, "dur": 2.050,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656022254.788, "dur": 0.633,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656022256.160, "dur": 0.516,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656022260.074, "dur": 0.428,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656022264.031, "dur": 0.369,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656022267.658, "dur": 0.328,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656022268.657, "dur": 3.804,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656022273.405, "dur": 0.294,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656022277.292, "dur": 0.305,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656022296.512, "dur": 43.659,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656022373.657, "dur": 117.504,
    "args": {
      "External id": 2478613,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656022387.328, "dur": 3.747,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656022396.208, "dur": 12.433,
    "args": {
      "External id": 2478615,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656022400.889, "dur": 7.359,
    "args": {
      "External id": 2478616,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1336753,
    "ts": 1514656022406.256, "dur": 0.762,
    "args": {
      "External id": 2478617,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656022415.133, "dur": 31.555,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656022416.833, "dur": 2.974,
    "args": {
      "External id": 2478619,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656022422.689, "dur": 0.237,
    "args": {
      "External id": 2478620,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1336753,
    "ts": 1514656022423.704, "dur": 0.229,
    "args": {
      "External id": 2478621,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1336753,
    "ts": 1514656022429.300, "dur": 1.210,
    "args": {
      "External id": 2478622,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1336753,
    "ts": 1514656022431.010, "dur": 0.416,
    "args": {
      "External id": 2478623,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1336753,
    "ts": 1514656022431.903, "dur": 0.403,
    "args": {
      "External id": 2478624,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1336753,
    "ts": 1514656022436.899, "dur": 0.148,
    "args": {
      "External id": 2478625,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1336753,
    "ts": 1514656022437.928, "dur": 0.201,
    "args": {
      "External id": 2478626,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1336753,
    "ts": 1514656022438.580, "dur": 2.399,
    "args": {
      "External id": 2478627,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1336753,
    "ts": 1514656022458.835, "dur": 24.213,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656022562.448, "dur": 487.760,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656022598.803, "dur": 440.826,
    "args": {
      "External id": 2478630,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1336753,
    "ts": 1514656022609.262, "dur": 417.982,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656023090.064, "dur": 4.880,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.4)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656023239.094, "dur": 16575.999,
    "args": {
      "External id": 2478633,"Record function id": 0, "Ev Idx": 6336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656023376.567, "dur": 10.006,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656023391.659, "dur": 1.970,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656023396.465, "dur": 2.804,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656023401.644, "dur": 1.248,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656023408.072, "dur": 1.617,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656023411.814, "dur": 1.074,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656023415.658, "dur": 1.722,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656023422.472, "dur": 3.278,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656023430.491, "dur": 1.095,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656023433.682, "dur": 1.173,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656023458.362, "dur": 16295.181,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656023482.125, "dur": 16260.326,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656023513.784, "dur": 39.626,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656023558.533, "dur": 16136.585,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656023561.258, "dur": 16132.681,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656023568.494, "dur": 8.536,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656023582.256, "dur": 16105.583,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656039969.589, "dur": 67.533,
    "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": 1336753, "tid": 1336753, "ts": 1514656039969.589,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336753, "tid": 1336753,
    "ts": 1514656040016.888, "dur": 14.706,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656040024.532, "dur": 6.715,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656040109.884, "dur": 81.752,
    "args": {
      "External id": 2478654,"Record function id": 0, "Ev Idx": 6357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336753, "tid": 1336753,
    "ts": 1514656040193.052, "dur": 1534.314,
    "args": {
      "External id": 2478655,"Record function id": 0, "Ev Idx": 6358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656040232.292, "dur": 1475.805,
    "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": 1336753, "tid": 1336753, "ts": 1514656040232.292,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336753, "tid": 1336753,
    "ts": 1514656040303.094, "dur": 48.622,
    "args": {
      "External id": 2478657,"kernel_hash": "cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/ml/cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc.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": 1336753, "tid": 1336753,
    "ts": 1514656040364.478, "dur": 105.939,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656040483.416, "dur": 81.781,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656040583.054, "dur": 48.286,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656040667.637, "dur": 37.623,
    "args": {
      "External id": 2478661,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656040731.115, "dur": 22.446,
    "args": {
      "External id": 2478662,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656040779.878, "dur": 174.405,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656040836.946, "dur": 13.902,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656040842.563, "dur": 7.200,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656040854.070, "dur": 7.525,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656040865.522, "dur": 1.254,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656040870.216, "dur": 5.330,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656040979.374, "dur": 159.072,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656041194.388, "dur": 51.595,
    "args": {
      "External id": 2478670,"kernel_hash": "cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/w7/cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5.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": 1336753, "tid": 1336753,
    "ts": 1514656041258.816, "dur": 66.669,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656041336.445, "dur": 56.193,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656041423.578, "dur": 36.135,
    "args": {
      "External id": 2478673,"kernel_hash": "czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/zh/czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha.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": 1336753, "tid": 1336753,
    "ts": 1514656041468.624, "dur": 49.021,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656041573.989, "dur": 30.187,
    "args": {
      "External id": 2478675,"kernel_hash": "cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/nz/cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4.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": 1336753, "tid": 1336753,
    "ts": 1514656041807.299, "dur": 92.702,
    "args": {
      "External id": 2478676,"Record function id": 0, "Ev Idx": 6379
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336753, "tid": 1336753,
    "ts": 1514656042049.011, "dur": 66.821,
    "args": {
      "External id": 2478677,"Record function id": 0, "Ev Idx": 6380
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.5)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656042129.750, "dur": 16746.562,
    "args": {
      "External id": 2478678,"Record function id": 0, "Ev Idx": 6381
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.5)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656042144.238, "dur": 1273.932,
    "args": {
      "External id": 2478679,"Record function id": 0, "Ev Idx": 6382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514656042243.522, "dur": 12.638,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656042273.697, "dur": 53.391,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656042283.019, "dur": 2.647,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656042290.963, "dur": 0.770,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656042295.608, "dur": 0.515,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656042297.688, "dur": 0.697,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656042302.721, "dur": 0.413,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656042306.488, "dur": 0.389,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656042308.226, "dur": 3.462,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656042313.004, "dur": 0.507,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656042316.972, "dur": 0.441,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656042342.001, "dur": 57.678,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656042443.555, "dur": 192.837,
    "args": {
      "External id": 2478692,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656042462.908, "dur": 5.382,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656042474.967, "dur": 12.170,
    "args": {
      "External id": 2478694,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656042480.091, "dur": 6.506,
    "args": {
      "External id": 2478695,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1336753,
    "ts": 1514656042483.833, "dur": 0.963,
    "args": {
      "External id": 2478696,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656042496.366, "dur": 65.974,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656042500.515, "dur": 2.646,
    "args": {
      "External id": 2478698,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656042504.692, "dur": 0.489,
    "args": {
      "External id": 2478699,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1336753,
    "ts": 1514656042506.471, "dur": 0.723,
    "args": {
      "External id": 2478700,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1336753,
    "ts": 1514656042513.317, "dur": 2.112,
    "args": {
      "External id": 2478701,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1336753,
    "ts": 1514656042516.733, "dur": 0.829,
    "args": {
      "External id": 2478702,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1336753,
    "ts": 1514656042518.954, "dur": 0.279,
    "args": {
      "External id": 2478703,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1336753,
    "ts": 1514656042525.388, "dur": 0.279,
    "args": {
      "External id": 2478704,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1336753,
    "ts": 1514656042546.791, "dur": 0.682,
    "args": {
      "External id": 2478705,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1336753,
    "ts": 1514656042551.707, "dur": 2.879,
    "args": {
      "External id": 2478706,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1336753,
    "ts": 1514656042583.263, "dur": 40.506,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656042708.717, "dur": 556.378,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656042747.476, "dur": 507.746,
    "args": {
      "External id": 2478709,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1336753,
    "ts": 1514656042761.734, "dur": 482.834,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656043306.209, "dur": 4.110,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.5)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656043447.403, "dur": 15219.446,
    "args": {
      "External id": 2478712,"Record function id": 0, "Ev Idx": 6415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656043597.547, "dur": 8.668,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656043611.054, "dur": 1.586,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656043615.164, "dur": 3.006,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656043620.350, "dur": 1.108,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656043623.043, "dur": 0.925,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656043625.124, "dur": 1.263,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656043630.845, "dur": 0.768,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656043633.015, "dur": 2.376,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656043636.657, "dur": 0.666,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656043638.905, "dur": 0.895,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656043661.784, "dur": 14951.700,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656043690.541, "dur": 14914.398,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656043713.780, "dur": 15.142,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656043732.072, "dur": 14833.629,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656043734.796, "dur": 14829.987,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656043740.610, "dur": 5.248,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656043747.671, "dur": 14813.484,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656058816.255, "dur": 33.578,
    "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": 1336753, "tid": 1336753, "ts": 1514656058816.255,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336753, "tid": 1336753,
    "ts": 1514656058832.912, "dur": 12.069,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656058839.098, "dur": 5.639,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656058917.457, "dur": 113.593,
    "args": {
      "External id": 2478733,"Record function id": 0, "Ev Idx": 6436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336753, "tid": 1336753,
    "ts": 1514656059034.338, "dur": 1140.630,
    "args": {
      "External id": 2478734,"Record function id": 0, "Ev Idx": 6437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656059082.152, "dur": 1079.301,
    "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": 1336753, "tid": 1336753, "ts": 1514656059082.152,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336753, "tid": 1336753,
    "ts": 1514656059152.905, "dur": 49.745,
    "args": {
      "External id": 2478736,"kernel_hash": "cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/ml/cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc.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": 1336753, "tid": 1336753,
    "ts": 1514656059215.417, "dur": 104.815,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656059330.952, "dur": 41.695,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656059380.842, "dur": 31.775,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656059438.423, "dur": 28.993,
    "args": {
      "External id": 2478740,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656059486.893, "dur": 15.383,
    "args": {
      "External id": 2478741,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656059549.938, "dur": 136.563,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656059601.932, "dur": 11.787,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656059606.772, "dur": 5.974,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656059617.917, "dur": 4.611,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656059623.553, "dur": 0.918,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656059626.655, "dur": 5.496,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656059697.797, "dur": 51.812,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656059782.618, "dur": 33.861,
    "args": {
      "External id": 2478749,"kernel_hash": "cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/w7/cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5.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": 1336753, "tid": 1336753,
    "ts": 1514656059825.407, "dur": 41.689,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656059877.550, "dur": 34.931,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656059933.797, "dur": 26.627,
    "args": {
      "External id": 2478752,"kernel_hash": "czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/zh/czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha.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": 1336753, "tid": 1336753,
    "ts": 1514656059966.409, "dur": 73.365,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656060066.808, "dur": 25.614,
    "args": {
      "External id": 2478754,"kernel_hash": "cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/nz/cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4.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": 1336753, "tid": 1336753,
    "ts": 1514656060240.198, "dur": 72.705,
    "args": {
      "External id": 2478755,"Record function id": 0, "Ev Idx": 6458
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336753, "tid": 1336753,
    "ts": 1514656060389.845, "dur": 46.910,
    "args": {
      "External id": 2478756,"Record function id": 0, "Ev Idx": 6459
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.6)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656060445.602, "dur": 18405.946,
    "args": {
      "External id": 2478757,"Record function id": 0, "Ev Idx": 6460
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.6)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656060459.074, "dur": 923.786,
    "args": {
      "External id": 2478758,"Record function id": 0, "Ev Idx": 6461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514656060557.362, "dur": 8.886,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656060580.600, "dur": 41.551,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656060585.811, "dur": 2.295,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656060592.593, "dur": 0.622,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656060595.115, "dur": 0.566,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656060596.674, "dur": 0.404,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656060601.340, "dur": 0.595,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656060603.836, "dur": 0.459,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656060606.010, "dur": 2.597,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656060610.205, "dur": 0.476,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656060612.623, "dur": 0.281,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656060634.888, "dur": 43.473,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656060713.528, "dur": 123.785,
    "args": {
      "External id": 2478771,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656060726.583, "dur": 4.370,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656060735.985, "dur": 10.685,
    "args": {
      "External id": 2478773,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656060740.549, "dur": 5.705,
    "args": {
      "External id": 2478774,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1336753,
    "ts": 1514656060744.351, "dur": 0.560,
    "args": {
      "External id": 2478775,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656060754.266, "dur": 37.961,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656060756.841, "dur": 2.657,
    "args": {
      "External id": 2478777,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656060762.220, "dur": 0.572,
    "args": {
      "External id": 2478778,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1336753,
    "ts": 1514656060764.088, "dur": 0.689,
    "args": {
      "External id": 2478779,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1336753,
    "ts": 1514656060769.532, "dur": 1.703,
    "args": {
      "External id": 2478780,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1336753,
    "ts": 1514656060773.152, "dur": 0.455,
    "args": {
      "External id": 2478781,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1336753,
    "ts": 1514656060775.054, "dur": 0.518,
    "args": {
      "External id": 2478782,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1336753,
    "ts": 1514656060779.403, "dur": 0.146,
    "args": {
      "External id": 2478783,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1336753,
    "ts": 1514656060781.715, "dur": 0.365,
    "args": {
      "External id": 2478784,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1336753,
    "ts": 1514656060783.727, "dur": 2.240,
    "args": {
      "External id": 2478785,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1336753,
    "ts": 1514656060804.274, "dur": 24.459,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656060889.496, "dur": 391.095,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656060923.891, "dur": 351.108,
    "args": {
      "External id": 2478788,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1336753,
    "ts": 1514656060935.957, "dur": 332.539,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656061311.112, "dur": 2.474,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.6)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656061402.476, "dur": 17250.606,
    "args": {
      "External id": 2478791,"Record function id": 0, "Ev Idx": 6494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656061499.591, "dur": 6.657,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656061510.542, "dur": 0.834,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656061513.409, "dur": 2.632,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656061518.009, "dur": 0.882,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656061520.749, "dur": 0.993,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656061523.458, "dur": 0.874,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656061545.459, "dur": 1.826,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656061550.537, "dur": 3.154,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656061555.347, "dur": 0.718,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656061557.709, "dur": 0.777,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656061580.329, "dur": 17022.075,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656061596.868, "dur": 16998.510,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656061619.837, "dur": 13.646,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656061636.218, "dur": 16923.770,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656061638.647, "dur": 16920.620,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656061645.083, "dur": 5.063,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656061651.967, "dur": 16904.593,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656078797.893, "dur": 29.467,
    "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": 1336753, "tid": 1336753, "ts": 1514656078797.893,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336753, "tid": 1336753,
    "ts": 1514656078813.883, "dur": 8.958,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656078817.958, "dur": 4.650,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656078893.859, "dur": 79.727,
    "args": {
      "External id": 2478812,"Record function id": 0, "Ev Idx": 6515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336753, "tid": 1336753,
    "ts": 1514656078975.467, "dur": 1130.383,
    "args": {
      "External id": 2478813,"Record function id": 0, "Ev Idx": 6516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656079050.042, "dur": 1041.986,
    "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": 1336753, "tid": 1336753, "ts": 1514656079050.042,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336753, "tid": 1336753,
    "ts": 1514656079117.365, "dur": 44.847,
    "args": {
      "External id": 2478815,"kernel_hash": "cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/ml/cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc.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": 1336753, "tid": 1336753,
    "ts": 1514656079174.721, "dur": 103.247,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656079289.564, "dur": 38.085,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656079335.930, "dur": 31.663,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656079393.512, "dur": 26.557,
    "args": {
      "External id": 2478819,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656079439.428, "dur": 17.612,
    "args": {
      "External id": 2478820,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656079476.336, "dur": 153.738,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656079542.586, "dur": 15.088,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656079548.610, "dur": 7.874,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656079561.237, "dur": 5.640,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656079568.311, "dur": 0.969,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656079572.301, "dur": 4.153,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656079641.500, "dur": 48.922,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656079722.537, "dur": 32.428,
    "args": {
      "External id": 2478828,"kernel_hash": "cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/w7/cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5.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": 1336753, "tid": 1336753,
    "ts": 1514656079764.577, "dur": 42.236,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656079814.925, "dur": 35.374,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656079876.206, "dur": 29.375,
    "args": {
      "External id": 2478831,"kernel_hash": "czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/zh/czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha.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": 1336753, "tid": 1336753,
    "ts": 1514656079911.191, "dur": 33.618,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656079964.724, "dur": 51.584,
    "args": {
      "External id": 2478833,"kernel_hash": "cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/nz/cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4.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": 1336753, "tid": 1336753,
    "ts": 1514656080172.400, "dur": 74.430,
    "args": {
      "External id": 2478834,"Record function id": 0, "Ev Idx": 6537
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336753, "tid": 1336753,
    "ts": 1514656080323.361, "dur": 48.136,
    "args": {
      "External id": 2478835,"Record function id": 0, "Ev Idx": 6538
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.7)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656080381.423, "dur": 18264.431,
    "args": {
      "External id": 2478836,"Record function id": 0, "Ev Idx": 6539
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.7)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656080390.587, "dur": 922.011,
    "args": {
      "External id": 2478837,"Record function id": 0, "Ev Idx": 6540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514656080471.412, "dur": 8.372,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656080494.038, "dur": 58.696,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656080499.592, "dur": 2.018,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656080506.388, "dur": 0.380,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656080509.018, "dur": 0.213,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656080510.584, "dur": 0.400,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656080515.030, "dur": 0.400,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656080516.872, "dur": 0.241,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656080518.812, "dur": 4.226,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656080524.477, "dur": 0.225,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656080542.508, "dur": 0.595,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656080566.421, "dur": 42.795,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656080646.277, "dur": 122.503,
    "args": {
      "External id": 2478850,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656080659.856, "dur": 3.995,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656080669.300, "dur": 10.206,
    "args": {
      "External id": 2478852,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656080673.512, "dur": 5.612,
    "args": {
      "External id": 2478853,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1336753,
    "ts": 1514656080677.293, "dur": 0.607,
    "args": {
      "External id": 2478854,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656080686.907, "dur": 35.727,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656080689.624, "dur": 2.640,
    "args": {
      "External id": 2478856,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656080693.880, "dur": 0.222,
    "args": {
      "External id": 2478857,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1336753,
    "ts": 1514656080695.727, "dur": 0.284,
    "args": {
      "External id": 2478858,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1336753,
    "ts": 1514656080700.200, "dur": 1.411,
    "args": {
      "External id": 2478859,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1336753,
    "ts": 1514656080703.588, "dur": 0.445,
    "args": {
      "External id": 2478860,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1336753,
    "ts": 1514656080705.715, "dur": 0.243,
    "args": {
      "External id": 2478861,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1336753,
    "ts": 1514656080709.682, "dur": 0.485,
    "args": {
      "External id": 2478862,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1336753,
    "ts": 1514656080711.739, "dur": 0.147,
    "args": {
      "External id": 2478863,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1336753,
    "ts": 1514656080714.121, "dur": 2.581,
    "args": {
      "External id": 2478864,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1336753,
    "ts": 1514656080736.415, "dur": 24.048,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656080820.934, "dur": 392.289,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656080857.200, "dur": 350.516,
    "args": {
      "External id": 2478867,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1336753,
    "ts": 1514656080869.502, "dur": 330.601,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656081240.775, "dur": 2.513,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.7)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656081332.820, "dur": 17098.433,
    "args": {
      "External id": 2478870,"Record function id": 0, "Ev Idx": 6573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656081431.916, "dur": 6.639,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656081443.069, "dur": 1.293,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656081446.435, "dur": 2.235,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656081452.611, "dur": 1.111,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656081455.486, "dur": 1.509,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656081458.764, "dur": 0.854,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656081461.353, "dur": 0.794,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656081466.933, "dur": 2.070,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656081470.637, "dur": 0.536,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656081472.875, "dur": 0.877,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656081494.422, "dur": 16888.252,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656081510.285, "dur": 16865.148,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656081548.279, "dur": 14.127,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656081567.925, "dur": 16772.761,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656081570.430, "dur": 16769.665,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656081576.777, "dur": 5.257,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656081583.811, "dur": 16753.516,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656098586.932, "dur": 32.049,
    "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": 1336753, "tid": 1336753, "ts": 1514656098586.932,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336753, "tid": 1336753,
    "ts": 1514656098605.850, "dur": 8.365,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656098609.232, "dur": 4.602,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656098687.275, "dur": 78.338,
    "args": {
      "External id": 2478891,"Record function id": 0, "Ev Idx": 6594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336753, "tid": 1336753,
    "ts": 1514656098768.117, "dur": 1087.967,
    "args": {
      "External id": 2478892,"Record function id": 0, "Ev Idx": 6595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656098810.070, "dur": 1032.941,
    "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": 1336753, "tid": 1336753, "ts": 1514656098810.070,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336753, "tid": 1336753,
    "ts": 1514656098879.448, "dur": 45.561,
    "args": {
      "External id": 2478894,"kernel_hash": "cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/ml/cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc.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": 1336753, "tid": 1336753,
    "ts": 1514656098937.452, "dur": 129.802,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656099079.797, "dur": 41.488,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656099130.526, "dur": 30.479,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656099188.401, "dur": 29.372,
    "args": {
      "External id": 2478898,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656099236.579, "dur": 15.705,
    "args": {
      "External id": 2478899,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656099272.684, "dur": 138.075,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656099321.013, "dur": 11.530,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656099326.469, "dur": 5.218,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656099335.372, "dur": 6.483,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656099351.051, "dur": 1.119,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656099354.911, "dur": 4.255,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656099421.040, "dur": 44.563,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656099498.603, "dur": 27.674,
    "args": {
      "External id": 2478907,"kernel_hash": "cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/w7/cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5.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": 1336753, "tid": 1336753,
    "ts": 1514656099552.553, "dur": 45.063,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656099605.999, "dur": 34.002,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656099663.533, "dur": 28.562,
    "args": {
      "External id": 2478910,"kernel_hash": "czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/zh/czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha.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": 1336753, "tid": 1336753,
    "ts": 1514656099697.163, "dur": 34.038,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656099752.384, "dur": 19.455,
    "args": {
      "External id": 2478912,"kernel_hash": "cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/nz/cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4.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": 1336753, "tid": 1336753,
    "ts": 1514656099918.977, "dur": 110.131,
    "args": {
      "External id": 2478913,"Record function id": 0, "Ev Idx": 6616
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336753, "tid": 1336753,
    "ts": 1514656100109.187, "dur": 49.678,
    "args": {
      "External id": 2478914,"Record function id": 0, "Ev Idx": 6617
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.8)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656100168.028, "dur": 18088.778,
    "args": {
      "External id": 2478915,"Record function id": 0, "Ev Idx": 6618
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.8)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656100175.788, "dur": 963.289,
    "args": {
      "External id": 2478916,"Record function id": 0, "Ev Idx": 6619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514656100257.653, "dur": 8.208,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656100280.336, "dur": 44.969,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656100286.797, "dur": 2.007,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656100294.113, "dur": 0.417,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656100296.219, "dur": 0.754,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656100298.949, "dur": 0.436,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656100303.243, "dur": 0.558,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656100306.174, "dur": 0.419,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656100308.764, "dur": 4.577,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656100315.452, "dur": 0.614,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656100317.730, "dur": 0.149,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656100336.623, "dur": 42.775,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656100413.111, "dur": 133.974,
    "args": {
      "External id": 2478929,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656100425.527, "dur": 4.488,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656100435.069, "dur": 10.562,
    "args": {
      "External id": 2478931,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656100439.623, "dur": 5.519,
    "args": {
      "External id": 2478932,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1336753,
    "ts": 1514656100443.384, "dur": 0.526,
    "args": {
      "External id": 2478933,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656100452.449, "dur": 34.363,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656100455.021, "dur": 2.592,
    "args": {
      "External id": 2478935,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656100459.765, "dur": 0.407,
    "args": {
      "External id": 2478936,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1336753,
    "ts": 1514656100462.045, "dur": 0.211,
    "args": {
      "External id": 2478937,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1336753,
    "ts": 1514656100466.585, "dur": 2.003,
    "args": {
      "External id": 2478938,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1336753,
    "ts": 1514656100470.396, "dur": 0.350,
    "args": {
      "External id": 2478939,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1336753,
    "ts": 1514656100472.521, "dur": 0.164,
    "args": {
      "External id": 2478940,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1336753,
    "ts": 1514656100475.112, "dur": 0.290,
    "args": {
      "External id": 2478941,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1336753,
    "ts": 1514656100477.050, "dur": 0.348,
    "args": {
      "External id": 2478942,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1336753,
    "ts": 1514656100478.844, "dur": 2.688,
    "args": {
      "External id": 2478943,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1336753,
    "ts": 1514656100498.444, "dur": 24.652,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656100602.653, "dur": 427.550,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656100639.018, "dur": 385.441,
    "args": {
      "External id": 2478946,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1336753,
    "ts": 1514656100650.563, "dur": 366.298,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656101060.139, "dur": 3.109,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.8)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656101160.347, "dur": 16897.811,
    "args": {
      "External id": 2478949,"Record function id": 0, "Ev Idx": 6652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656101259.013, "dur": 6.879,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656101270.308, "dur": 1.013,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656101273.280, "dur": 2.830,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656101278.075, "dur": 0.993,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656101280.585, "dur": 1.116,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656101283.177, "dur": 1.254,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656101287.941, "dur": 0.980,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656101290.440, "dur": 2.513,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656101294.502, "dur": 0.895,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656101297.091, "dur": 0.593,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656101328.932, "dur": 16682.636,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656101349.053, "dur": 16655.155,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656101369.212, "dur": 14.047,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656101386.060, "dur": 16562.842,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656101388.502, "dur": 16559.916,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656101394.146, "dur": 5.123,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656101400.807, "dur": 16544.822,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656118202.834, "dur": 28.348,
    "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": 1336753, "tid": 1336753, "ts": 1514656118202.834,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336753, "tid": 1336753,
    "ts": 1514656118219.084, "dur": 7.571,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656118222.344, "dur": 4.076,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656118298.333, "dur": 73.655,
    "args": {
      "External id": 2478970,"Record function id": 0, "Ev Idx": 6673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336753, "tid": 1336753,
    "ts": 1514656118373.895, "dur": 1091.854,
    "args": {
      "External id": 2478971,"Record function id": 0, "Ev Idx": 6674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656118415.216, "dur": 1037.084,
    "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": 1336753, "tid": 1336753, "ts": 1514656118415.216,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336753, "tid": 1336753,
    "ts": 1514656118481.041, "dur": 41.685,
    "args": {
      "External id": 2478973,"kernel_hash": "cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/ml/cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc.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": 1336753, "tid": 1336753,
    "ts": 1514656118552.718, "dur": 103.003,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656118666.136, "dur": 37.614,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656118712.675, "dur": 30.748,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656118769.815, "dur": 27.819,
    "args": {
      "External id": 2478977,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656118815.972, "dur": 17.187,
    "args": {
      "External id": 2478978,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656118851.222, "dur": 165.768,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656118900.115, "dur": 10.824,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656118904.921, "dur": 5.311,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656118914.353, "dur": 6.640,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656118922.350, "dur": 0.923,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656118925.902, "dur": 4.354,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656119031.454, "dur": 52.370,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656119115.987, "dur": 29.625,
    "args": {
      "External id": 2478986,"kernel_hash": "cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/w7/cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5.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": 1336753, "tid": 1336753,
    "ts": 1514656119153.987, "dur": 41.938,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656119204.302, "dur": 34.334,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656119263.160, "dur": 29.568,
    "args": {
      "External id": 2478989,"kernel_hash": "czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/zh/czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha.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": 1336753, "tid": 1336753,
    "ts": 1514656119298.485, "dur": 34.246,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656119352.243, "dur": 23.606,
    "args": {
      "External id": 2478991,"kernel_hash": "cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/nz/cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4.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": 1336753, "tid": 1336753,
    "ts": 1514656119543.023, "dur": 75.756,
    "args": {
      "External id": 2478992,"Record function id": 0, "Ev Idx": 6695
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336753, "tid": 1336753,
    "ts": 1514656119698.794, "dur": 46.883,
    "args": {
      "External id": 2478993,"Record function id": 0, "Ev Idx": 6696
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.9)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656119755.005, "dur": 17996.205,
    "args": {
      "External id": 2478994,"Record function id": 0, "Ev Idx": 6697
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.9)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656119763.658, "dur": 966.019,
    "args": {
      "External id": 2478995,"Record function id": 0, "Ev Idx": 6698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514656119845.161, "dur": 8.151,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656119867.488, "dur": 41.785,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656119873.493, "dur": 2.205,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656119880.272, "dur": 0.471,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656119882.645, "dur": 0.392,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656119884.134, "dur": 0.474,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656119888.070, "dur": 0.376,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656119889.585, "dur": 0.434,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656119891.368, "dur": 4.416,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656119897.524, "dur": 0.329,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656119899.735, "dur": 0.742,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656119920.313, "dur": 43.035,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656120036.707, "dur": 163.340,
    "args": {
      "External id": 2479008,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656120050.652, "dur": 5.098,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656120061.159, "dur": 10.744,
    "args": {
      "External id": 2479010,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656120065.398, "dur": 6.088,
    "args": {
      "External id": 2479011,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1336753,
    "ts": 1514656120068.908, "dur": 0.777,
    "args": {
      "External id": 2479012,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656120078.916, "dur": 35.463,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656120080.760, "dur": 2.748,
    "args": {
      "External id": 2479014,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656120087.910, "dur": 0.363,
    "args": {
      "External id": 2479015,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1336753,
    "ts": 1514656120089.529, "dur": 0.367,
    "args": {
      "External id": 2479016,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1336753,
    "ts": 1514656120093.006, "dur": 2.396,
    "args": {
      "External id": 2479017,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1336753,
    "ts": 1514656120096.781, "dur": 0.311,
    "args": {
      "External id": 2479018,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1336753,
    "ts": 1514656120098.314, "dur": 2.445,
    "args": {
      "External id": 2479019,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1336753,
    "ts": 1514656120101.916, "dur": 0.408,
    "args": {
      "External id": 2479020,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1336753,
    "ts": 1514656120103.716, "dur": 0.469,
    "args": {
      "External id": 2479021,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1336753,
    "ts": 1514656120107.868, "dur": 0.412,
    "args": {
      "External id": 2479022,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1336753,
    "ts": 1514656120160.388, "dur": 30.391,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656120255.233, "dur": 374.896,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656120288.021, "dur": 336.777,
    "args": {
      "External id": 2479025,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1336753,
    "ts": 1514656120302.025, "dur": 316.617,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656120658.327, "dur": 2.253,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.9)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656120749.616, "dur": 16809.150,
    "args": {
      "External id": 2479028,"Record function id": 0, "Ev Idx": 6731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656120847.477, "dur": 6.695,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656120857.972, "dur": 0.987,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656120860.811, "dur": 2.494,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656120864.998, "dur": 1.080,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656120867.444, "dur": 0.802,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656120869.652, "dur": 0.805,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656120874.475, "dur": 0.848,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656120876.709, "dur": 3.076,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656120881.430, "dur": 0.995,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656120883.588, "dur": 1.035,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656120903.857, "dur": 16595.257,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656120919.229, "dur": 16573.413,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656120940.732, "dur": 13.144,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656120956.931, "dur": 16501.938,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656120959.454, "dur": 16498.763,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656120966.029, "dur": 5.904,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656120973.463, "dur": 16482.161,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656137698.659, "dur": 27.967,
    "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": 1336753, "tid": 1336753, "ts": 1514656137698.659,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336753, "tid": 1336753,
    "ts": 1514656137714.185, "dur": 7.876,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656137717.258, "dur": 4.559,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656137791.689, "dur": 82.582,
    "args": {
      "External id": 2479049,"Record function id": 0, "Ev Idx": 6752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336753, "tid": 1336753,
    "ts": 1514656137875.950, "dur": 1081.370,
    "args": {
      "External id": 2479050,"Record function id": 0, "Ev Idx": 6753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656137915.194, "dur": 1028.674,
    "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": 1336753, "tid": 1336753, "ts": 1514656137915.194,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336753, "tid": 1336753,
    "ts": 1514656138011.773, "dur": 48.315,
    "args": {
      "External id": 2479052,"kernel_hash": "cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/ml/cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc.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": 1336753, "tid": 1336753,
    "ts": 1514656138075.868, "dur": 103.523,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656138188.225, "dur": 38.280,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656138235.464, "dur": 30.129,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656138293.317, "dur": 29.591,
    "args": {
      "External id": 2479056,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656138343.228, "dur": 16.328,
    "args": {
      "External id": 2479057,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656138378.935, "dur": 126.188,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656138426.213, "dur": 10.630,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656138431.163, "dur": 4.906,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656138439.417, "dur": 5.938,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656138446.761, "dur": 1.286,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656138450.389, "dur": 4.784,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656138516.192, "dur": 65.277,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656138614.690, "dur": 31.436,
    "args": {
      "External id": 2479065,"kernel_hash": "cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/w7/cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5.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": 1336753, "tid": 1336753,
    "ts": 1514656138662.905, "dur": 43.141,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656138713.483, "dur": 35.048,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656138773.042, "dur": 28.043,
    "args": {
      "External id": 2479068,"kernel_hash": "czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/zh/czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha.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": 1336753, "tid": 1336753,
    "ts": 1514656138806.681, "dur": 33.332,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656138858.881, "dur": 19.373,
    "args": {
      "External id": 2479070,"kernel_hash": "cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/nz/cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4.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": 1336753, "tid": 1336753,
    "ts": 1514656139058.466, "dur": 78.793,
    "args": {
      "External id": 2479071,"Record function id": 0, "Ev Idx": 6774
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336753, "tid": 1336753,
    "ts": 1514656139214.467, "dur": 49.636,
    "args": {
      "External id": 2479072,"Record function id": 0, "Ev Idx": 6775
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.10)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656139273.211, "dur": 18235.773,
    "args": {
      "External id": 2479073,"Record function id": 0, "Ev Idx": 6776
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.10)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656139280.514, "dur": 1004.238,
    "args": {
      "External id": 2479074,"Record function id": 0, "Ev Idx": 6777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514656139361.206, "dur": 7.898,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656139383.300, "dur": 38.078,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656139388.682, "dur": 2.279,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656139395.351, "dur": 0.337,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656139396.778, "dur": 0.462,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656139398.523, "dur": 0.499,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656139402.424, "dur": 0.331,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656139404.242, "dur": 0.554,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656139406.180, "dur": 4.255,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656139411.861, "dur": 0.168,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656139413.004, "dur": 0.423,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656139433.210, "dur": 41.486,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656139507.503, "dur": 135.235,
    "args": {
      "External id": 2479087,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656139520.421, "dur": 4.125,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656139546.795, "dur": 10.911,
    "args": {
      "External id": 2479089,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656139551.275, "dur": 6.017,
    "args": {
      "External id": 2479090,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1336753,
    "ts": 1514656139554.790, "dur": 0.738,
    "args": {
      "External id": 2479091,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656139565.150, "dur": 30.336,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656139567.146, "dur": 2.722,
    "args": {
      "External id": 2479093,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656139570.881, "dur": 0.552,
    "args": {
      "External id": 2479094,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1336753,
    "ts": 1514656139572.660, "dur": 0.376,
    "args": {
      "External id": 2479095,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1336753,
    "ts": 1514656139576.622, "dur": 2.054,
    "args": {
      "External id": 2479096,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1336753,
    "ts": 1514656139579.740, "dur": 0.435,
    "args": {
      "External id": 2479097,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1336753,
    "ts": 1514656139581.221, "dur": 0.266,
    "args": {
      "External id": 2479098,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1336753,
    "ts": 1514656139584.778, "dur": 0.180,
    "args": {
      "External id": 2479099,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1336753,
    "ts": 1514656139586.528, "dur": 0.383,
    "args": {
      "External id": 2479100,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1336753,
    "ts": 1514656139588.096, "dur": 2.240,
    "args": {
      "External id": 2479101,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1336753,
    "ts": 1514656139607.631, "dur": 26.829,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656139697.147, "dur": 482.527,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656139730.499, "dur": 442.937,
    "args": {
      "External id": 2479104,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1336753,
    "ts": 1514656139832.087, "dur": 335.201,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656140207.713, "dur": 2.699,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.10)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656140305.156, "dur": 16981.509,
    "args": {
      "External id": 2479107,"Record function id": 0, "Ev Idx": 6810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656140401.620, "dur": 6.805,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656140412.371, "dur": 1.040,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656140415.348, "dur": 3.074,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656140420.663, "dur": 0.710,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656140422.977, "dur": 0.841,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656140425.071, "dur": 0.718,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656140429.536, "dur": 0.849,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656140431.903, "dur": 2.112,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656140435.764, "dur": 0.862,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656140437.957, "dur": 0.669,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656140458.841, "dur": 16756.627,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656140474.163, "dur": 16730.305,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656140492.323, "dur": 14.190,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656140509.428, "dur": 16649.422,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656140511.818, "dur": 16645.948,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656140517.641, "dur": 5.992,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656140525.322, "dur": 16626.963,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656157448.175, "dur": 33.729,
    "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": 1336753, "tid": 1336753, "ts": 1514656157448.175,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336753, "tid": 1336753,
    "ts": 1514656157465.239, "dur": 11.800,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656157470.437, "dur": 6.353,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656157565.874, "dur": 80.880,
    "args": {
      "External id": 2479128,"Record function id": 0, "Ev Idx": 6831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336753, "tid": 1336753,
    "ts": 1514656157648.922, "dur": 1114.283,
    "args": {
      "External id": 2479129,"Record function id": 0, "Ev Idx": 6832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656157691.591, "dur": 1057.785,
    "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": 1336753, "tid": 1336753, "ts": 1514656157691.591,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336753, "tid": 1336753,
    "ts": 1514656157763.143, "dur": 47.183,
    "args": {
      "External id": 2479131,"kernel_hash": "cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/ml/cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc.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": 1336753, "tid": 1336753,
    "ts": 1514656157823.926, "dur": 102.847,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656157936.857, "dur": 39.073,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656158021.513, "dur": 38.152,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656158091.243, "dur": 26.665,
    "args": {
      "External id": 2479135,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656158136.212, "dur": 18.238,
    "args": {
      "External id": 2479136,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656158175.067, "dur": 134.441,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656158227.172, "dur": 11.314,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656158232.671, "dur": 5.047,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656158241.764, "dur": 6.331,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656158249.515, "dur": 0.936,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656158252.854, "dur": 5.460,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656158320.557, "dur": 46.416,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656158396.850, "dur": 31.212,
    "args": {
      "External id": 2479144,"kernel_hash": "cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/w7/cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5.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": 1336753, "tid": 1336753,
    "ts": 1514656158437.500, "dur": 41.986,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656158487.521, "dur": 34.653,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656158561.835, "dur": 33.159,
    "args": {
      "External id": 2479147,"kernel_hash": "czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/zh/czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha.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": 1336753, "tid": 1336753,
    "ts": 1514656158601.660, "dur": 37.235,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656158661.013, "dur": 19.363,
    "args": {
      "External id": 2479149,"kernel_hash": "cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/nz/cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4.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": 1336753, "tid": 1336753,
    "ts": 1514656158826.382, "dur": 76.877,
    "args": {
      "External id": 2479150,"Record function id": 0, "Ev Idx": 6853
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336753, "tid": 1336753,
    "ts": 1514656158978.688, "dur": 87.911,
    "args": {
      "External id": 2479151,"Record function id": 0, "Ev Idx": 6854
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.11)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656159077.877, "dur": 17889.734,
    "args": {
      "External id": 2479152,"Record function id": 0, "Ev Idx": 6855
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.11)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656159086.439, "dur": 875.061,
    "args": {
      "External id": 2479153,"Record function id": 0, "Ev Idx": 6856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514656159168.934, "dur": 9.676,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656159193.516, "dur": 40.495,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656159199.162, "dur": 1.997,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656159205.941, "dur": 0.228,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656159207.715, "dur": 0.665,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656159210.428, "dur": 0.163,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656159214.825, "dur": 0.548,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656159216.614, "dur": 0.429,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656159218.666, "dur": 4.211,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656159224.197, "dur": 0.165,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656159226.177, "dur": 0.313,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656159249.024, "dur": 45.983,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656159330.682, "dur": 118.862,
    "args": {
      "External id": 2479166,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656159343.537, "dur": 5.602,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656159354.475, "dur": 10.643,
    "args": {
      "External id": 2479168,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656159358.810, "dur": 5.893,
    "args": {
      "External id": 2479169,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1336753,
    "ts": 1514656159362.847, "dur": 0.618,
    "args": {
      "External id": 2479170,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656159372.182, "dur": 32.631,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656159374.899, "dur": 0.620,
    "args": {
      "External id": 2479172,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656159377.645, "dur": 2.949,
    "args": {
      "External id": 2479173,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1336753,
    "ts": 1514656159382.028, "dur": 0.360,
    "args": {
      "External id": 2479174,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1336753,
    "ts": 1514656159383.949, "dur": 1.666,
    "args": {
      "External id": 2479175,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1336753,
    "ts": 1514656159389.182, "dur": 0.317,
    "args": {
      "External id": 2479176,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1336753,
    "ts": 1514656159390.948, "dur": 0.413,
    "args": {
      "External id": 2479177,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1336753,
    "ts": 1514656159392.953, "dur": 0.401,
    "args": {
      "External id": 2479178,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1336753,
    "ts": 1514656159397.308, "dur": 0.284,
    "args": {
      "External id": 2479179,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1336753,
    "ts": 1514656159398.989, "dur": 0.536,
    "args": {
      "External id": 2479180,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1336753,
    "ts": 1514656159417.105, "dur": 24.900,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656159500.293, "dur": 365.092,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656159549.113, "dur": 311.112,
    "args": {
      "External id": 2479183,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1336753,
    "ts": 1514656159562.621, "dur": 292.338,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656159891.760, "dur": 2.133,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.11)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656160017.465, "dur": 16750.849,
    "args": {
      "External id": 2479186,"Record function id": 0, "Ev Idx": 6889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656160120.867, "dur": 7.139,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656160131.678, "dur": 1.354,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656160134.873, "dur": 2.456,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656160138.851, "dur": 0.554,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656160140.751, "dur": 0.696,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656160142.749, "dur": 0.661,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656160147.711, "dur": 0.740,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656160150.031, "dur": 2.047,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656160153.476, "dur": 0.982,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656160155.886, "dur": 0.924,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656160178.223, "dur": 16551.074,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656160194.385, "dur": 16528.483,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656160215.041, "dur": 13.345,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656160231.140, "dur": 16457.440,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656160233.672, "dur": 16454.425,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656160240.246, "dur": 6.588,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656160248.478, "dur": 16437.100,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656176904.622, "dur": 38.731,
    "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": 1336753, "tid": 1336753, "ts": 1514656176904.622,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336753, "tid": 1336753,
    "ts": 1514656176931.208, "dur": 7.616,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656176934.431, "dur": 4.164,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656177036.903, "dur": 81.944,
    "args": {
      "External id": 2479207,"Record function id": 0, "Ev Idx": 6910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336753, "tid": 1336753,
    "ts": 1514656177120.981, "dur": 1094.011,
    "args": {
      "External id": 2479208,"Record function id": 0, "Ev Idx": 6911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656177162.682, "dur": 1039.725,
    "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": 1336753, "tid": 1336753, "ts": 1514656177162.682,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336753, "tid": 1336753,
    "ts": 1514656177230.778, "dur": 44.739,
    "args": {
      "External id": 2479210,"kernel_hash": "cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/ml/cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc.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": 1336753, "tid": 1336753,
    "ts": 1514656177288.457, "dur": 105.614,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656177403.928, "dur": 38.584,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656177450.339, "dur": 30.253,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656177507.367, "dur": 43.835,
    "args": {
      "External id": 2479214,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656177572.724, "dur": 18.086,
    "args": {
      "External id": 2479215,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656177611.114, "dur": 130.181,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656177660.302, "dur": 11.568,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656177665.471, "dur": 5.584,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656177674.778, "dur": 5.714,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656177682.191, "dur": 1.000,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656177685.534, "dur": 5.556,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656177752.143, "dur": 50.302,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656177833.853, "dur": 29.519,
    "args": {
      "External id": 2479223,"kernel_hash": "cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/w7/cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5.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": 1336753, "tid": 1336753,
    "ts": 1514656177872.422, "dur": 40.602,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656177922.505, "dur": 34.696,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656177977.240, "dur": 66.050,
    "args": {
      "External id": 2479226,"kernel_hash": "czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/zh/czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha.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": 1336753, "tid": 1336753,
    "ts": 1514656178051.177, "dur": 39.022,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656178114.284, "dur": 18.485,
    "args": {
      "External id": 2479228,"kernel_hash": "cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/nz/cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4.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": 1336753, "tid": 1336753,
    "ts": 1514656178279.699, "dur": 75.943,
    "args": {
      "External id": 2479229,"Record function id": 0, "Ev Idx": 6932
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336753, "tid": 1336753,
    "ts": 1514656178431.506, "dur": 48.239,
    "args": {
      "External id": 2479230,"Record function id": 0, "Ev Idx": 6933
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.12)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656178488.641, "dur": 18141.751,
    "args": {
      "External id": 2479231,"Record function id": 0, "Ev Idx": 6934
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.12)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656178496.153, "dur": 908.880,
    "args": {
      "External id": 2479232,"Record function id": 0, "Ev Idx": 6935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514656178595.439, "dur": 8.662,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656178621.694, "dur": 41.910,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656178627.546, "dur": 2.254,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656178635.083, "dur": 0.652,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656178637.229, "dur": 0.435,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656178639.031, "dur": 2.964,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656178643.680, "dur": 0.304,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656178645.471, "dur": 0.566,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656178650.255, "dur": 2.284,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656178654.101, "dur": 0.396,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656178656.300, "dur": 0.184,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656178675.737, "dur": 42.418,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656178752.490, "dur": 118.195,
    "args": {
      "External id": 2479245,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656178765.489, "dur": 3.875,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656178774.163, "dur": 12.760,
    "args": {
      "External id": 2479247,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656178778.539, "dur": 7.971,
    "args": {
      "External id": 2479248,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1336753,
    "ts": 1514656178782.453, "dur": 2.972,
    "args": {
      "External id": 2479249,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656178794.070, "dur": 32.987,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656178797.497, "dur": 0.378,
    "args": {
      "External id": 2479251,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656178799.599, "dur": 0.628,
    "args": {
      "External id": 2479252,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1336753,
    "ts": 1514656178802.067, "dur": 0.522,
    "args": {
      "External id": 2479253,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1336753,
    "ts": 1514656178806.447, "dur": 1.746,
    "args": {
      "External id": 2479254,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1336753,
    "ts": 1514656178809.803, "dur": 0.257,
    "args": {
      "External id": 2479255,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1336753,
    "ts": 1514656178811.525, "dur": 3.008,
    "args": {
      "External id": 2479256,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1336753,
    "ts": 1514656178816.234, "dur": 0.267,
    "args": {
      "External id": 2479257,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1336753,
    "ts": 1514656178817.866, "dur": 0.169,
    "args": {
      "External id": 2479258,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1336753,
    "ts": 1514656178821.558, "dur": 0.559,
    "args": {
      "External id": 2479259,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1336753,
    "ts": 1514656178838.069, "dur": 24.978,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656178922.166, "dur": 383.070,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656178954.276, "dur": 345.545,
    "args": {
      "External id": 2479262,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1336753,
    "ts": 1514656178966.734, "dur": 326.454,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656179332.977, "dur": 2.532,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.12)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656179425.144, "dur": 16980.840,
    "args": {
      "External id": 2479265,"Record function id": 0, "Ev Idx": 6968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656179520.832, "dur": 24.589,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656179551.419, "dur": 1.851,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656179554.964, "dur": 3.163,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656179559.707, "dur": 1.124,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656179562.154, "dur": 1.111,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656179564.451, "dur": 0.987,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656179569.287, "dur": 0.782,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656179571.603, "dur": 2.165,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656179575.229, "dur": 0.649,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656179577.576, "dur": 0.804,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656179601.100, "dur": 16760.111,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656179617.663, "dur": 16736.048,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656179641.264, "dur": 14.515,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656179658.800, "dur": 16657.877,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656179661.285, "dur": 16654.678,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656179667.341, "dur": 6.245,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656179675.455, "dur": 16636.910,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656196563.581, "dur": 38.774,
    "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": 1336753, "tid": 1336753, "ts": 1514656196563.581,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336753, "tid": 1336753,
    "ts": 1514656196587.605, "dur": 9.858,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656196591.941, "dur": 5.110,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656196672.141, "dur": 78.433,
    "args": {
      "External id": 2479286,"Record function id": 0, "Ev Idx": 6989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336753, "tid": 1336753,
    "ts": 1514656196752.340, "dur": 1093.106,
    "args": {
      "External id": 2479287,"Record function id": 0, "Ev Idx": 6990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656196793.922, "dur": 1038.731,
    "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": 1336753, "tid": 1336753, "ts": 1514656196793.922,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336753, "tid": 1336753,
    "ts": 1514656196862.084, "dur": 44.885,
    "args": {
      "External id": 2479289,"kernel_hash": "cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/ml/cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc.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": 1336753, "tid": 1336753,
    "ts": 1514656196920.036, "dur": 132.474,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656197066.545, "dur": 42.491,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656197117.437, "dur": 31.371,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656197176.685, "dur": 29.885,
    "args": {
      "External id": 2479293,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656197225.461, "dur": 18.999,
    "args": {
      "External id": 2479294,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656197263.784, "dur": 131.284,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656197313.678, "dur": 11.048,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656197318.714, "dur": 4.985,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656197327.947, "dur": 5.653,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656197334.893, "dur": 1.114,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656197338.554, "dur": 6.011,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656197405.079, "dur": 43.796,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656197479.001, "dur": 28.263,
    "args": {
      "External id": 2479302,"kernel_hash": "cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/w7/cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5.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": 1336753, "tid": 1336753,
    "ts": 1514656197516.519, "dur": 57.504,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656197585.809, "dur": 36.543,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656197651.911, "dur": 31.947,
    "args": {
      "External id": 2479305,"kernel_hash": "czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/zh/czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha.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": 1336753, "tid": 1336753,
    "ts": 1514656197691.569, "dur": 33.739,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656197744.388, "dur": 20.051,
    "args": {
      "External id": 2479307,"kernel_hash": "cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/nz/cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4.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": 1336753, "tid": 1336753,
    "ts": 1514656197907.366, "dur": 110.457,
    "args": {
      "External id": 2479308,"Record function id": 0, "Ev Idx": 7011
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336753, "tid": 1336753,
    "ts": 1514656198100.348, "dur": 48.985,
    "args": {
      "External id": 2479309,"Record function id": 0, "Ev Idx": 7012
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.13)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656198158.273, "dur": 18190.468,
    "args": {
      "External id": 2479310,"Record function id": 0, "Ev Idx": 7013
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.13)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656198166.421, "dur": 950.085,
    "args": {
      "External id": 2479311,"Record function id": 0, "Ev Idx": 7014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514656198248.217, "dur": 9.055,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656198271.648, "dur": 41.471,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656198277.916, "dur": 2.445,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656198285.340, "dur": 0.499,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656198287.396, "dur": 0.529,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656198289.575, "dur": 0.270,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656198293.867, "dur": 0.692,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656198295.929, "dur": 0.264,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656198297.961, "dur": 4.678,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656198304.313, "dur": 0.354,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656198306.318, "dur": 0.168,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656198325.538, "dur": 43.097,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656198402.435, "dur": 121.328,
    "args": {
      "External id": 2479324,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656198416.035, "dur": 3.980,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656198425.020, "dur": 10.403,
    "args": {
      "External id": 2479326,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656198429.514, "dur": 5.496,
    "args": {
      "External id": 2479327,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1336753,
    "ts": 1514656198433.336, "dur": 0.437,
    "args": {
      "External id": 2479328,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656198442.546, "dur": 34.922,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656198445.266, "dur": 2.753,
    "args": {
      "External id": 2479330,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656198449.712, "dur": 0.237,
    "args": {
      "External id": 2479331,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1336753,
    "ts": 1514656198451.413, "dur": 0.948,
    "args": {
      "External id": 2479332,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1336753,
    "ts": 1514656198456.379, "dur": 1.840,
    "args": {
      "External id": 2479333,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1336753,
    "ts": 1514656198459.779, "dur": 0.592,
    "args": {
      "External id": 2479334,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1336753,
    "ts": 1514656198462.318, "dur": 0.314,
    "args": {
      "External id": 2479335,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1336753,
    "ts": 1514656198466.271, "dur": 0.367,
    "args": {
      "External id": 2479336,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1336753,
    "ts": 1514656198468.307, "dur": 0.150,
    "args": {
      "External id": 2479337,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1336753,
    "ts": 1514656198470.031, "dur": 2.622,
    "args": {
      "External id": 2479338,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1336753,
    "ts": 1514656198490.653, "dur": 25.622,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656198593.423, "dur": 380.746,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656198628.256, "dur": 340.882,
    "args": {
      "External id": 2479341,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1336753,
    "ts": 1514656198638.430, "dur": 324.737,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656199037.576, "dur": 3.260,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.13)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656199138.268, "dur": 17013.492,
    "args": {
      "External id": 2479344,"Record function id": 0, "Ev Idx": 7047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656199242.433, "dur": 6.790,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656199252.600, "dur": 1.331,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656199255.673, "dur": 2.180,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656199259.586, "dur": 1.269,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656199262.008, "dur": 1.057,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656199264.326, "dur": 0.872,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656199268.986, "dur": 0.916,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656199271.730, "dur": 1.815,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656199275.238, "dur": 0.692,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656199277.661, "dur": 0.877,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656199300.021, "dur": 16812.125,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656199317.175, "dur": 16788.114,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656199343.027, "dur": 13.829,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656199359.754, "dur": 16711.975,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656199362.499, "dur": 16708.626,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656199368.793, "dur": 4.951,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656199375.420, "dur": 16692.912,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656216286.868, "dur": 35.794,
    "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": 1336753, "tid": 1336753, "ts": 1514656216286.868,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336753, "tid": 1336753,
    "ts": 1514656216310.507, "dur": 7.493,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656216313.683, "dur": 4.153,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656216389.801, "dur": 77.252,
    "args": {
      "External id": 2479365,"Record function id": 0, "Ev Idx": 7068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336753, "tid": 1336753,
    "ts": 1514656216468.425, "dur": 1120.482,
    "args": {
      "External id": 2479366,"Record function id": 0, "Ev Idx": 7069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656216508.163, "dur": 1066.557,
    "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": 1336753, "tid": 1336753, "ts": 1514656216508.163,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336753, "tid": 1336753,
    "ts": 1514656216595.601, "dur": 46.635,
    "args": {
      "External id": 2479368,"kernel_hash": "cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/ml/cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc.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": 1336753, "tid": 1336753,
    "ts": 1514656216654.909, "dur": 103.407,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656216769.282, "dur": 38.590,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656216816.105, "dur": 31.875,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656216872.809, "dur": 29.350,
    "args": {
      "External id": 2479372,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656216921.498, "dur": 15.422,
    "args": {
      "External id": 2479373,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656216957.298, "dur": 175.677,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656217049.099, "dur": 12.046,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656217054.522, "dur": 5.636,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656217064.080, "dur": 5.699,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656217071.057, "dur": 1.096,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656217074.432, "dur": 4.655,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656217145.040, "dur": 52.174,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656217229.638, "dur": 30.885,
    "args": {
      "External id": 2479381,"kernel_hash": "cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/w7/cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5.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": 1336753, "tid": 1336753,
    "ts": 1514656217269.207, "dur": 41.259,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656217318.580, "dur": 35.164,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656217378.334, "dur": 27.335,
    "args": {
      "External id": 2479384,"kernel_hash": "czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/zh/czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha.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": 1336753, "tid": 1336753,
    "ts": 1514656217411.681, "dur": 34.328,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656217466.816, "dur": 20.882,
    "args": {
      "External id": 2479386,"kernel_hash": "cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/nz/cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4.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": 1336753, "tid": 1336753,
    "ts": 1514656217653.247, "dur": 72.301,
    "args": {
      "External id": 2479387,"Record function id": 0, "Ev Idx": 7090
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336753, "tid": 1336753,
    "ts": 1514656217801.708, "dur": 47.730,
    "args": {
      "External id": 2479388,"Record function id": 0, "Ev Idx": 7091
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.14)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656217860.109, "dur": 18069.555,
    "args": {
      "External id": 2479389,"Record function id": 0, "Ev Idx": 7092
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.14)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656217868.555, "dur": 937.386,
    "args": {
      "External id": 2479390,"Record function id": 0, "Ev Idx": 7093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514656217946.356, "dur": 8.164,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656217969.361, "dur": 87.612,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656217980.533, "dur": 38.771,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656218025.398, "dur": 0.434,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656218027.350, "dur": 0.647,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656218029.681, "dur": 3.243,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656218034.428, "dur": 0.691,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656218038.991, "dur": 0.620,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656218043.258, "dur": 2.111,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656218046.770, "dur": 0.527,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656218048.658, "dur": 0.887,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656218070.380, "dur": 46.734,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656218153.320, "dur": 120.841,
    "args": {
      "External id": 2479403,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656218167.387, "dur": 5.166,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656218178.062, "dur": 10.131,
    "args": {
      "External id": 2479405,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656218182.165, "dur": 5.583,
    "args": {
      "External id": 2479406,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1336753,
    "ts": 1514656218185.871, "dur": 0.626,
    "args": {
      "External id": 2479407,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656218196.197, "dur": 33.126,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656218199.377, "dur": 0.347,
    "args": {
      "External id": 2479409,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656218201.567, "dur": 0.581,
    "args": {
      "External id": 2479410,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1336753,
    "ts": 1514656218203.648, "dur": 2.302,
    "args": {
      "External id": 2479411,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1336753,
    "ts": 1514656218207.660, "dur": 2.036,
    "args": {
      "External id": 2479412,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1336753,
    "ts": 1514656218211.350, "dur": 0.395,
    "args": {
      "External id": 2479413,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1336753,
    "ts": 1514656218215.696, "dur": 0.300,
    "args": {
      "External id": 2479414,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1336753,
    "ts": 1514656218217.357, "dur": 0.359,
    "args": {
      "External id": 2479415,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1336753,
    "ts": 1514656218219.309, "dur": 0.360,
    "args": {
      "External id": 2479416,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1336753,
    "ts": 1514656218223.224, "dur": 0.173,
    "args": {
      "External id": 2479417,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1336753,
    "ts": 1514656218240.950, "dur": 25.197,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656218326.474, "dur": 380.490,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656218361.414, "dur": 340.483,
    "args": {
      "External id": 2479420,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1336753,
    "ts": 1514656218371.726, "dur": 324.113,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656218734.480, "dur": 2.011,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.14)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656218826.695, "dur": 16907.486,
    "args": {
      "External id": 2479423,"Record function id": 0, "Ev Idx": 7126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656218923.759, "dur": 6.595,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656218933.784, "dur": 1.558,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656218937.099, "dur": 2.988,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656218941.765, "dur": 1.604,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656218944.933, "dur": 1.102,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656218947.192, "dur": 0.749,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656218952.046, "dur": 0.614,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656218953.969, "dur": 2.557,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656218958.029, "dur": 0.970,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656218960.977, "dur": 0.674,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656219016.816, "dur": 16677.972,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656219034.886, "dur": 16653.162,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656219061.143, "dur": 14.788,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656219078.680, "dur": 16577.460,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656219081.363, "dur": 16574.264,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656219087.739, "dur": 6.742,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656219096.358, "dur": 16556.577,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656235868.851, "dur": 37.206,
    "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": 1336753, "tid": 1336753, "ts": 1514656235868.851,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336753, "tid": 1336753,
    "ts": 1514656235893.547, "dur": 7.610,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656235896.658, "dur": 4.329,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656235970.958, "dur": 106.648,
    "args": {
      "External id": 2479444,"Record function id": 0, "Ev Idx": 7147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336753, "tid": 1336753,
    "ts": 1514656236080.861, "dur": 1102.867,
    "args": {
      "External id": 2479445,"Record function id": 0, "Ev Idx": 7148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656236121.831, "dur": 1048.062,
    "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": 1336753, "tid": 1336753, "ts": 1514656236121.831,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336753, "tid": 1336753,
    "ts": 1514656236189.777, "dur": 44.046,
    "args": {
      "External id": 2479447,"kernel_hash": "cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/ml/cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc.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": 1336753, "tid": 1336753,
    "ts": 1514656236246.948, "dur": 102.248,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656236359.286, "dur": 39.426,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656236407.452, "dur": 31.733,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656236464.610, "dur": 28.176,
    "args": {
      "External id": 2479451,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656236513.194, "dur": 32.227,
    "args": {
      "External id": 2479452,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656236567.938, "dur": 135.406,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656236622.073, "dur": 11.433,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656236627.638, "dur": 5.000,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656236636.678, "dur": 5.230,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656236643.214, "dur": 0.899,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656236647.019, "dur": 4.543,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656236713.909, "dur": 50.255,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656236798.031, "dur": 31.372,
    "args": {
      "External id": 2479460,"kernel_hash": "cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/w7/cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5.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": 1336753, "tid": 1336753,
    "ts": 1514656236839.564, "dur": 42.721,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656236890.687, "dur": 34.435,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656236945.974, "dur": 28.320,
    "args": {
      "External id": 2479463,"kernel_hash": "czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/zh/czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha.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": 1336753, "tid": 1336753,
    "ts": 1514656236979.605, "dur": 74.590,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656237080.589, "dur": 21.219,
    "args": {
      "External id": 2479465,"kernel_hash": "cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/nz/cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4.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": 1336753, "tid": 1336753,
    "ts": 1514656237247.244, "dur": 77.562,
    "args": {
      "External id": 2479466,"Record function id": 0, "Ev Idx": 7169
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336753, "tid": 1336753,
    "ts": 1514656237401.758, "dur": 48.371,
    "args": {
      "External id": 2479467,"Record function id": 0, "Ev Idx": 7170
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.15)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656237459.822, "dur": 18103.092,
    "args": {
      "External id": 2479468,"Record function id": 0, "Ev Idx": 7171
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.15)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656237468.198, "dur": 944.598,
    "args": {
      "External id": 2479469,"Record function id": 0, "Ev Idx": 7172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514656237568.110, "dur": 9.166,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656237591.967, "dur": 40.744,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656237597.829, "dur": 2.096,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656237604.630, "dur": 0.391,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656237606.840, "dur": 0.231,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656237608.925, "dur": 0.540,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656237612.792, "dur": 0.339,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656237614.544, "dur": 0.379,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656237616.547, "dur": 4.591,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656237622.723, "dur": 0.320,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656237624.520, "dur": 0.398,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656237645.144, "dur": 41.418,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656237721.000, "dur": 158.622,
    "args": {
      "External id": 2479482,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656237731.786, "dur": 3.658,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656237740.778, "dur": 10.609,
    "args": {
      "External id": 2479484,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656237745.110, "dur": 5.856,
    "args": {
      "External id": 2479485,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1336753,
    "ts": 1514656237748.952, "dur": 0.786,
    "args": {
      "External id": 2479486,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656237758.880, "dur": 34.319,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656237761.852, "dur": 2.532,
    "args": {
      "External id": 2479488,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656237766.247, "dur": 0.372,
    "args": {
      "External id": 2479489,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1336753,
    "ts": 1514656237768.135, "dur": 0.577,
    "args": {
      "External id": 2479490,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1336753,
    "ts": 1514656237773.030, "dur": 2.561,
    "args": {
      "External id": 2479491,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1336753,
    "ts": 1514656237777.107, "dur": 0.259,
    "args": {
      "External id": 2479492,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1336753,
    "ts": 1514656237778.946, "dur": 0.359,
    "args": {
      "External id": 2479493,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1336753,
    "ts": 1514656237783.070, "dur": 0.406,
    "args": {
      "External id": 2479494,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1336753,
    "ts": 1514656237785.036, "dur": 0.206,
    "args": {
      "External id": 2479495,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1336753,
    "ts": 1514656237786.647, "dur": 2.047,
    "args": {
      "External id": 2479496,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1336753,
    "ts": 1514656237844.040, "dur": 27.398,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656237934.113, "dur": 380.170,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656237968.700, "dur": 340.192,
    "args": {
      "External id": 2479499,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1336753,
    "ts": 1514656238014.834, "dur": 285.815,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656238342.609, "dur": 2.109,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.15)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656238433.406, "dur": 16917.599,
    "args": {
      "External id": 2479502,"Record function id": 0, "Ev Idx": 7205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656238547.185, "dur": 6.659,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656238558.618, "dur": 1.181,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656238561.635, "dur": 3.103,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656238566.634, "dur": 0.792,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656238568.869, "dur": 0.905,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656238571.025, "dur": 1.025,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656238575.932, "dur": 0.778,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656238578.392, "dur": 1.945,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656238582.014, "dur": 0.644,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656238584.431, "dur": 0.691,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656238606.569, "dur": 16706.920,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656238623.746, "dur": 16682.728,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656238645.099, "dur": 15.759,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656238663.768, "dur": 16609.077,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656238666.474, "dur": 16605.825,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656238672.666, "dur": 6.193,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656238680.414, "dur": 16589.361,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656255487.893, "dur": 36.405,
    "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": 1336753, "tid": 1336753, "ts": 1514656255487.893,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336753, "tid": 1336753,
    "ts": 1514656255511.898, "dur": 7.788,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656255515.293, "dur": 4.244,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656255606.109, "dur": 73.878,
    "args": {
      "External id": 2479523,"Record function id": 0, "Ev Idx": 7226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336753, "tid": 1336753,
    "ts": 1514656255681.525, "dur": 1092.608,
    "args": {
      "External id": 2479524,"Record function id": 0, "Ev Idx": 7227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656255726.776, "dur": 1035.055,
    "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": 1336753, "tid": 1336753, "ts": 1514656255726.776,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336753, "tid": 1336753,
    "ts": 1514656255790.826, "dur": 45.158,
    "args": {
      "External id": 2479526,"kernel_hash": "cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/ml/cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc.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": 1336753, "tid": 1336753,
    "ts": 1514656255847.403, "dur": 99.873,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656255957.419, "dur": 74.441,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656256045.161, "dur": 35.938,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656256110.022, "dur": 28.322,
    "args": {
      "External id": 2479530,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656256156.981, "dur": 16.162,
    "args": {
      "External id": 2479531,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656256193.159, "dur": 131.628,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656256243.753, "dur": 11.870,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656256249.241, "dur": 5.389,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656256258.712, "dur": 5.340,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656256265.636, "dur": 1.300,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656256269.457, "dur": 4.186,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656256335.188, "dur": 46.816,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656256413.074, "dur": 29.287,
    "args": {
      "External id": 2479539,"kernel_hash": "cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/w7/cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5.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": 1336753, "tid": 1336753,
    "ts": 1514656256450.830, "dur": 40.830,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656256500.090, "dur": 49.497,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656256575.865, "dur": 28.213,
    "args": {
      "External id": 2479542,"kernel_hash": "czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/zh/czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha.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": 1336753, "tid": 1336753,
    "ts": 1514656256610.168, "dur": 37.355,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656256667.958, "dur": 20.308,
    "args": {
      "External id": 2479544,"kernel_hash": "cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/nz/cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4.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": 1336753, "tid": 1336753,
    "ts": 1514656256836.160, "dur": 74.771,
    "args": {
      "External id": 2479545,"Record function id": 0, "Ev Idx": 7248
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336753, "tid": 1336753,
    "ts": 1514656257022.592, "dur": 50.425,
    "args": {
      "External id": 2479546,"Record function id": 0, "Ev Idx": 7249
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.16)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656257084.218, "dur": 18104.754,
    "args": {
      "External id": 2479547,"Record function id": 0, "Ev Idx": 7250
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.16)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656257093.449, "dur": 866.971,
    "args": {
      "External id": 2479548,"Record function id": 0, "Ev Idx": 7251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514656257173.572, "dur": 8.572,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656257197.377, "dur": 39.292,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656257203.548, "dur": 2.196,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656257210.730, "dur": 0.450,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656257212.419, "dur": 0.260,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656257214.509, "dur": 0.312,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656257218.374, "dur": 0.602,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656257220.597, "dur": 0.360,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656257222.428, "dur": 3.635,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656257227.579, "dur": 0.396,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656257229.405, "dur": 0.161,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656257252.748, "dur": 42.289,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656257329.025, "dur": 117.101,
    "args": {
      "External id": 2479561,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656257339.362, "dur": 6.337,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656257350.729, "dur": 10.591,
    "args": {
      "External id": 2479563,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656257355.123, "dur": 5.794,
    "args": {
      "External id": 2479564,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1336753,
    "ts": 1514656257359.168, "dur": 0.528,
    "args": {
      "External id": 2479565,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656257368.918, "dur": 32.656,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656257371.725, "dur": 0.507,
    "args": {
      "External id": 2479567,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656257373.932, "dur": 2.590,
    "args": {
      "External id": 2479568,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1336753,
    "ts": 1514656257378.114, "dur": 0.466,
    "args": {
      "External id": 2479569,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1336753,
    "ts": 1514656257380.534, "dur": 2.074,
    "args": {
      "External id": 2479570,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1336753,
    "ts": 1514656257386.782, "dur": 0.433,
    "args": {
      "External id": 2479571,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1336753,
    "ts": 1514656257388.614, "dur": 0.485,
    "args": {
      "External id": 2479572,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1336753,
    "ts": 1514656257390.615, "dur": 0.557,
    "args": {
      "External id": 2479573,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1336753,
    "ts": 1514656257394.816, "dur": 0.333,
    "args": {
      "External id": 2479574,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1336753,
    "ts": 1514656257396.612, "dur": 0.355,
    "args": {
      "External id": 2479575,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1336753,
    "ts": 1514656257412.742, "dur": 25.237,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656257495.506, "dur": 365.050,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656257545.010, "dur": 310.319,
    "args": {
      "External id": 2479578,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1336753,
    "ts": 1514656257557.273, "dur": 292.509,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656257886.649, "dur": 2.094,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.16)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656258015.714, "dur": 16946.045,
    "args": {
      "External id": 2479581,"Record function id": 0, "Ev Idx": 7284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656258123.215, "dur": 7.208,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656258134.050, "dur": 1.657,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656258137.449, "dur": 2.262,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656258141.337, "dur": 1.243,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656258143.875, "dur": 1.101,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656258146.280, "dur": 1.282,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656258151.252, "dur": 1.018,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656258153.761, "dur": 2.539,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656258157.817, "dur": 0.966,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656258160.189, "dur": 0.938,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656258181.926, "dur": 16740.039,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656258197.928, "dur": 16717.500,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656258224.717, "dur": 14.538,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656258242.206, "dur": 16640.718,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656258244.884, "dur": 16637.457,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656258251.065, "dur": 5.916,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656258258.671, "dur": 16620.853,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656275127.400, "dur": 36.163,
    "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": 1336753, "tid": 1336753, "ts": 1514656275127.400,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336753, "tid": 1336753,
    "ts": 1514656275151.200, "dur": 7.555,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656275154.416, "dur": 4.056,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656275230.254, "dur": 78.581,
    "args": {
      "External id": 2479602,"Record function id": 0, "Ev Idx": 7305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336753, "tid": 1336753,
    "ts": 1514656275310.403, "dur": 1180.157,
    "args": {
      "External id": 2479603,"Record function id": 0, "Ev Idx": 7306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656275352.052, "dur": 1124.391,
    "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": 1336753, "tid": 1336753, "ts": 1514656275352.052,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336753, "tid": 1336753,
    "ts": 1514656275416.899, "dur": 43.813,
    "args": {
      "External id": 2479605,"kernel_hash": "cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/ml/cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc.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": 1336753, "tid": 1336753,
    "ts": 1514656275472.847, "dur": 120.874,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656275606.701, "dur": 40.347,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656275656.001, "dur": 31.006,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656275721.648, "dur": 29.931,
    "args": {
      "External id": 2479609,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656275768.267, "dur": 16.770,
    "args": {
      "External id": 2479610,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656275804.169, "dur": 134.521,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656275854.265, "dur": 12.204,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656275859.858, "dur": 5.510,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656275869.849, "dur": 6.406,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656275877.660, "dur": 1.444,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656275881.554, "dur": 6.254,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656276033.468, "dur": 61.889,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656276138.181, "dur": 34.330,
    "args": {
      "External id": 2479618,"kernel_hash": "cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/w7/cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5.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": 1336753, "tid": 1336753,
    "ts": 1514656276183.602, "dur": 42.854,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656276235.417, "dur": 34.463,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656276292.820, "dur": 30.464,
    "args": {
      "External id": 2479621,"kernel_hash": "czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/zh/czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha.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": 1336753, "tid": 1336753,
    "ts": 1514656276328.541, "dur": 36.263,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656276387.430, "dur": 19.111,
    "args": {
      "External id": 2479623,"kernel_hash": "cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/nz/cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4.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": 1336753, "tid": 1336753,
    "ts": 1514656276573.367, "dur": 78.099,
    "args": {
      "External id": 2479624,"Record function id": 0, "Ev Idx": 7327
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336753, "tid": 1336753,
    "ts": 1514656276729.308, "dur": 46.591,
    "args": {
      "External id": 2479625,"Record function id": 0, "Ev Idx": 7328
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.17)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656276785.198, "dur": 18092.898,
    "args": {
      "External id": 2479626,"Record function id": 0, "Ev Idx": 7329
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.17)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656276793.275, "dur": 922.044,
    "args": {
      "External id": 2479627,"Record function id": 0, "Ev Idx": 7330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514656276872.602, "dur": 8.978,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656276895.885, "dur": 42.203,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656276902.247, "dur": 2.426,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656276909.729, "dur": 0.397,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656276911.905, "dur": 0.658,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656276914.721, "dur": 0.295,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656276919.152, "dur": 0.652,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656276921.543, "dur": 0.647,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656276923.804, "dur": 2.870,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656276928.770, "dur": 0.471,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656276930.731, "dur": 0.639,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656276950.667, "dur": 79.204,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656277067.782, "dur": 125.603,
    "args": {
      "External id": 2479640,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656277079.136, "dur": 5.100,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656277089.438, "dur": 10.994,
    "args": {
      "External id": 2479642,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656277093.799, "dur": 6.253,
    "args": {
      "External id": 2479643,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1336753,
    "ts": 1514656277097.789, "dur": 0.755,
    "args": {
      "External id": 2479644,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656277107.901, "dur": 33.654,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656277111.158, "dur": 2.719,
    "args": {
      "External id": 2479646,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656277115.697, "dur": 0.416,
    "args": {
      "External id": 2479647,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1336753,
    "ts": 1514656277117.586, "dur": 0.468,
    "args": {
      "External id": 2479648,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1336753,
    "ts": 1514656277122.112, "dur": 1.624,
    "args": {
      "External id": 2479649,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1336753,
    "ts": 1514656277125.384, "dur": 0.439,
    "args": {
      "External id": 2479650,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1336753,
    "ts": 1514656277127.197, "dur": 0.350,
    "args": {
      "External id": 2479651,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1336753,
    "ts": 1514656277130.933, "dur": 0.692,
    "args": {
      "External id": 2479652,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1336753,
    "ts": 1514656277133.095, "dur": 0.188,
    "args": {
      "External id": 2479653,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1336753,
    "ts": 1514656277134.926, "dur": 2.472,
    "args": {
      "External id": 2479654,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1336753,
    "ts": 1514656277156.327, "dur": 28.676,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656277248.415, "dur": 364.187,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656277283.611, "dur": 323.774,
    "args": {
      "External id": 2479657,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1336753,
    "ts": 1514656277293.534, "dur": 308.200,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656277640.643, "dur": 2.325,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.17)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656277736.476, "dur": 16929.408,
    "args": {
      "External id": 2479660,"Record function id": 0, "Ev Idx": 7363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656277838.082, "dur": 6.574,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656277847.823, "dur": 1.334,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656277850.983, "dur": 2.877,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656277855.337, "dur": 1.349,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656277858.100, "dur": 0.845,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656277860.057, "dur": 0.909,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656277864.587, "dur": 0.990,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656277866.963, "dur": 2.138,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656277870.639, "dur": 1.151,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656277873.474, "dur": 0.916,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656277895.149, "dur": 16723.015,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656277911.621, "dur": 16697.946,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656277941.403, "dur": 13.756,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656277957.903, "dur": 16617.075,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656277960.482, "dur": 16613.656,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656277966.718, "dur": 5.561,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656277974.029, "dur": 16596.985,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656294814.056, "dur": 37.867,
    "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": 1336753, "tid": 1336753, "ts": 1514656294814.056,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336753, "tid": 1336753,
    "ts": 1514656294837.529, "dur": 9.741,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656294841.913, "dur": 5.124,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656294919.934, "dur": 105.906,
    "args": {
      "External id": 2479681,"Record function id": 0, "Ev Idx": 7384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336753, "tid": 1336753,
    "ts": 1514656295029.321, "dur": 1108.079,
    "args": {
      "External id": 2479682,"Record function id": 0, "Ev Idx": 7385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656295072.983, "dur": 1050.156,
    "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": 1336753, "tid": 1336753, "ts": 1514656295072.983,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336753, "tid": 1336753,
    "ts": 1514656295144.919, "dur": 48.862,
    "args": {
      "External id": 2479684,"kernel_hash": "cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/ml/cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc.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": 1336753, "tid": 1336753,
    "ts": 1514656295206.411, "dur": 100.280,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656295316.850, "dur": 38.132,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656295363.964, "dur": 32.461,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656295424.029, "dur": 28.623,
    "args": {
      "External id": 2479688,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656295470.646, "dur": 16.553,
    "args": {
      "External id": 2479689,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656295506.433, "dur": 149.331,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656295572.896, "dur": 12.882,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656295578.467, "dur": 6.307,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656295588.659, "dur": 6.208,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656295596.145, "dur": 1.864,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656295600.421, "dur": 4.271,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656295666.737, "dur": 51.254,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656295750.473, "dur": 28.064,
    "args": {
      "External id": 2479697,"kernel_hash": "cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/w7/cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5.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": 1336753, "tid": 1336753,
    "ts": 1514656295787.243, "dur": 41.107,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656295837.135, "dur": 35.082,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656295897.310, "dur": 29.075,
    "args": {
      "External id": 2479700,"kernel_hash": "czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/zh/czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha.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": 1336753, "tid": 1336753,
    "ts": 1514656295931.913, "dur": 34.396,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656296024.026, "dur": 23.505,
    "args": {
      "External id": 2479702,"kernel_hash": "cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/nz/cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4.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": 1336753, "tid": 1336753,
    "ts": 1514656296204.104, "dur": 75.379,
    "args": {
      "External id": 2479703,"Record function id": 0, "Ev Idx": 7406
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336753, "tid": 1336753,
    "ts": 1514656296354.161, "dur": 47.002,
    "args": {
      "External id": 2479704,"Record function id": 0, "Ev Idx": 7407
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.18)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656296410.702, "dur": 18222.631,
    "args": {
      "External id": 2479705,"Record function id": 0, "Ev Idx": 7408
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.18)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656296419.380, "dur": 957.265,
    "args": {
      "External id": 2479706,"Record function id": 0, "Ev Idx": 7409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514656296498.419, "dur": 8.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656296519.928, "dur": 55.529,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656296525.635, "dur": 17.807,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656296549.870, "dur": 0.723,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656296552.043, "dur": 0.631,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656296554.426, "dur": 0.481,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656296558.371, "dur": 0.355,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656296560.516, "dur": 0.379,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656296562.256, "dur": 2.727,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656296566.743, "dur": 0.352,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656296568.296, "dur": 0.435,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656296588.651, "dur": 42.232,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656296665.833, "dur": 123.611,
    "args": {
      "External id": 2479719,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656296680.774, "dur": 7.524,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656296693.797, "dur": 10.331,
    "args": {
      "External id": 2479721,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656296698.295, "dur": 5.458,
    "args": {
      "External id": 2479722,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1336753,
    "ts": 1514656296701.818, "dur": 0.557,
    "args": {
      "External id": 2479723,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656296711.586, "dur": 33.472,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656296714.032, "dur": 0.540,
    "args": {
      "External id": 2479725,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656296716.401, "dur": 2.830,
    "args": {
      "External id": 2479726,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1336753,
    "ts": 1514656296720.842, "dur": 0.360,
    "args": {
      "External id": 2479727,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1336753,
    "ts": 1514656296723.391, "dur": 2.253,
    "args": {
      "External id": 2479728,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1336753,
    "ts": 1514656296729.815, "dur": 0.325,
    "args": {
      "External id": 2479729,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1336753,
    "ts": 1514656296731.729, "dur": 0.163,
    "args": {
      "External id": 2479730,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1336753,
    "ts": 1514656296733.412, "dur": 0.565,
    "args": {
      "External id": 2479731,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1336753,
    "ts": 1514656296737.982, "dur": 0.575,
    "args": {
      "External id": 2479732,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1336753,
    "ts": 1514656296740.323, "dur": 0.287,
    "args": {
      "External id": 2479733,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1336753,
    "ts": 1514656296755.565, "dur": 25.124,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656296841.454, "dur": 428.165,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656296876.190, "dur": 388.055,
    "args": {
      "External id": 2479736,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1336753,
    "ts": 1514656296886.481, "dur": 371.214,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656297300.087, "dur": 2.396,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.18)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656297398.279, "dur": 17027.278,
    "args": {
      "External id": 2479739,"Record function id": 0, "Ev Idx": 7442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656297497.157, "dur": 6.664,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656297507.025, "dur": 1.692,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656297510.459, "dur": 2.746,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656297514.777, "dur": 1.138,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656297517.339, "dur": 1.113,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656297519.622, "dur": 0.966,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656297524.655, "dur": 1.090,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656297545.653, "dur": 2.969,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656297551.389, "dur": 1.056,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656297554.047, "dur": 1.044,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656297577.420, "dur": 16810.396,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656297593.838, "dur": 16787.534,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656297619.592, "dur": 14.106,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656297636.434, "dur": 16711.347,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656297638.980, "dur": 16708.237,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656297645.410, "dur": 6.576,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656297653.599, "dur": 16690.823,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656314572.212, "dur": 35.688,
    "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": 1336753, "tid": 1336753, "ts": 1514656314572.212,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336753, "tid": 1336753,
    "ts": 1514656314594.965, "dur": 8.025,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656314598.275, "dur": 4.382,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656314674.256, "dur": 78.539,
    "args": {
      "External id": 2479760,"Record function id": 0, "Ev Idx": 7463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336753, "tid": 1336753,
    "ts": 1514656314754.267, "dur": 1082.581,
    "args": {
      "External id": 2479761,"Record function id": 0, "Ev Idx": 7464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656314795.127, "dur": 1028.526,
    "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": 1336753, "tid": 1336753, "ts": 1514656314795.127,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336753, "tid": 1336753,
    "ts": 1514656314869.002, "dur": 43.377,
    "args": {
      "External id": 2479763,"kernel_hash": "cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/ml/cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc.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": 1336753, "tid": 1336753,
    "ts": 1514656314926.994, "dur": 129.848,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656315070.860, "dur": 41.196,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656315118.479, "dur": 31.214,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656315179.283, "dur": 26.830,
    "args": {
      "External id": 2479767,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656315222.532, "dur": 20.351,
    "args": {
      "External id": 2479768,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656315260.502, "dur": 133.612,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656315310.457, "dur": 12.106,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656315316.090, "dur": 5.754,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656315325.312, "dur": 6.587,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656315333.197, "dur": 3.531,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656315339.257, "dur": 4.744,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656315404.207, "dur": 44.354,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656315479.301, "dur": 28.201,
    "args": {
      "External id": 2479776,"kernel_hash": "cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/w7/cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5.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": 1336753, "tid": 1336753,
    "ts": 1514656315516.676, "dur": 59.924,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656315587.093, "dur": 36.559,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656315651.156, "dur": 27.781,
    "args": {
      "External id": 2479779,"kernel_hash": "czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/zh/czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha.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": 1336753, "tid": 1336753,
    "ts": 1514656315684.309, "dur": 32.989,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656315735.171, "dur": 21.110,
    "args": {
      "External id": 2479781,"kernel_hash": "cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/nz/cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4.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": 1336753, "tid": 1336753,
    "ts": 1514656315900.762, "dur": 72.141,
    "args": {
      "External id": 2479782,"Record function id": 0, "Ev Idx": 7485
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336753, "tid": 1336753,
    "ts": 1514656316091.416, "dur": 51.453,
    "args": {
      "External id": 2479783,"Record function id": 0, "Ev Idx": 7486
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.19)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656316153.589, "dur": 18186.511,
    "args": {
      "External id": 2479784,"Record function id": 0, "Ev Idx": 7487
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.19)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656316162.587, "dur": 907.676,
    "args": {
      "External id": 2479785,"Record function id": 0, "Ev Idx": 7488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514656316244.806, "dur": 8.594,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656316268.803, "dur": 40.347,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656316274.928, "dur": 2.178,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656316282.168, "dur": 0.392,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656316283.996, "dur": 0.423,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656316286.190, "dur": 0.175,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656316289.839, "dur": 0.693,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656316292.079, "dur": 0.436,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656316293.811, "dur": 4.437,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656316299.979, "dur": 0.309,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656316302.004, "dur": 0.466,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656316320.388, "dur": 43.360,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656316397.542, "dur": 115.838,
    "args": {
      "External id": 2479798,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656316408.494, "dur": 4.505,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656316417.691, "dur": 10.486,
    "args": {
      "External id": 2479800,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656316422.258, "dur": 5.503,
    "args": {
      "External id": 2479801,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1336753,
    "ts": 1514656316426.165, "dur": 0.422,
    "args": {
      "External id": 2479802,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656316435.718, "dur": 32.297,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656316438.129, "dur": 2.621,
    "args": {
      "External id": 2479804,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656316442.529, "dur": 0.475,
    "args": {
      "External id": 2479805,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1336753,
    "ts": 1514656316444.931, "dur": 0.543,
    "args": {
      "External id": 2479806,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1336753,
    "ts": 1514656316449.457, "dur": 1.606,
    "args": {
      "External id": 2479807,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1336753,
    "ts": 1514656316452.554, "dur": 0.466,
    "args": {
      "External id": 2479808,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1336753,
    "ts": 1514656316454.424, "dur": 0.346,
    "args": {
      "External id": 2479809,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1336753,
    "ts": 1514656316458.531, "dur": 0.352,
    "args": {
      "External id": 2479810,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1336753,
    "ts": 1514656316460.271, "dur": 0.159,
    "args": {
      "External id": 2479811,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1336753,
    "ts": 1514656316461.809, "dur": 2.334,
    "args": {
      "External id": 2479812,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1336753,
    "ts": 1514656316480.620, "dur": 25.001,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656316583.367, "dur": 343.546,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656316618.189, "dur": 304.309,
    "args": {
      "External id": 2479815,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1336753,
    "ts": 1514656316628.726, "dur": 287.806,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656316951.943, "dur": 2.343,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.19)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656317094.415, "dur": 17042.210,
    "args": {
      "External id": 2479818,"Record function id": 0, "Ev Idx": 7521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656317196.965, "dur": 6.808,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656317233.797, "dur": 1.229,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656317237.141, "dur": 2.164,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656317242.809, "dur": 0.970,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656317245.178, "dur": 0.824,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656317247.131, "dur": 0.835,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656317249.704, "dur": 0.873,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656317254.228, "dur": 1.658,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656317257.519, "dur": 1.010,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656317260.206, "dur": 1.075,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656317282.320, "dur": 16813.561,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656317299.090, "dur": 16790.158,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656317318.397, "dur": 13.136,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656317337.140, "dur": 16718.183,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656317339.656, "dur": 16715.103,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656317345.694, "dur": 5.512,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656317352.853, "dur": 16699.242,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656334276.549, "dur": 39.024,
    "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": 1336753, "tid": 1336753, "ts": 1514656334276.549,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336753, "tid": 1336753,
    "ts": 1514656334302.608, "dur": 7.778,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656334305.823, "dur": 4.379,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656334381.305, "dur": 75.222,
    "args": {
      "External id": 2479839,"Record function id": 0, "Ev Idx": 7542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336753, "tid": 1336753,
    "ts": 1514656334458.339, "dur": 1106.061,
    "args": {
      "External id": 2479840,"Record function id": 0, "Ev Idx": 7543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656334496.591, "dur": 1053.421,
    "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": 1336753, "tid": 1336753, "ts": 1514656334496.591,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336753, "tid": 1336753,
    "ts": 1514656334580.090, "dur": 44.144,
    "args": {
      "External id": 2479842,"kernel_hash": "cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/ml/cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc.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": 1336753, "tid": 1336753,
    "ts": 1514656334636.859, "dur": 106.516,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656334753.193, "dur": 38.858,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656334800.953, "dur": 30.999,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656334858.669, "dur": 27.742,
    "args": {
      "External id": 2479846,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656334905.294, "dur": 16.628,
    "args": {
      "External id": 2479847,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656334940.020, "dur": 169.879,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656335023.846, "dur": 13.510,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656335029.484, "dur": 6.864,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656335040.571, "dur": 5.965,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656335048.310, "dur": 1.212,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656335051.958, "dur": 4.899,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656335121.441, "dur": 50.095,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656335205.009, "dur": 31.260,
    "args": {
      "External id": 2479855,"kernel_hash": "cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/w7/cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5.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": 1336753, "tid": 1336753,
    "ts": 1514656335245.496, "dur": 41.436,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656335295.751, "dur": 34.402,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656335354.309, "dur": 28.541,
    "args": {
      "External id": 2479858,"kernel_hash": "czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/zh/czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha.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": 1336753, "tid": 1336753,
    "ts": 1514656335388.575, "dur": 34.355,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656335443.812, "dur": 19.548,
    "args": {
      "External id": 2479860,"kernel_hash": "cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/nz/cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4.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": 1336753, "tid": 1336753,
    "ts": 1514656335626.723, "dur": 74.068,
    "args": {
      "External id": 2479861,"Record function id": 0, "Ev Idx": 7564
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336753, "tid": 1336753,
    "ts": 1514656335776.042, "dur": 49.396,
    "args": {
      "External id": 2479862,"Record function id": 0, "Ev Idx": 7565
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.20)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656335835.912, "dur": 18383.949,
    "args": {
      "External id": 2479863,"Record function id": 0, "Ev Idx": 7566
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.20)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656335845.128, "dur": 932.898,
    "args": {
      "External id": 2479864,"Record function id": 0, "Ev Idx": 7567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514656335924.550, "dur": 8.314,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656335947.535, "dur": 79.142,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656335953.676, "dur": 2.153,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656335960.808, "dur": 0.412,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656335962.893, "dur": 0.554,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656335965.440, "dur": 0.807,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656335969.385, "dur": 0.564,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656335971.540, "dur": 0.875,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656335973.695, "dur": 4.056,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656335979.533, "dur": 0.580,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656336019.124, "dur": 0.549,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656336041.179, "dur": 44.834,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656336122.559, "dur": 126.682,
    "args": {
      "External id": 2479877,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656336134.525, "dur": 6.165,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656336146.474, "dur": 10.873,
    "args": {
      "External id": 2479879,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656336150.859, "dur": 6.111,
    "args": {
      "External id": 2479880,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1336753,
    "ts": 1514656336154.973, "dur": 0.739,
    "args": {
      "External id": 2479881,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656336165.134, "dur": 39.237,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656336167.881, "dur": 3.247,
    "args": {
      "External id": 2479883,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656336172.853, "dur": 0.525,
    "args": {
      "External id": 2479884,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1336753,
    "ts": 1514656336178.831, "dur": 0.242,
    "args": {
      "External id": 2479885,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1336753,
    "ts": 1514656336183.085, "dur": 2.291,
    "args": {
      "External id": 2479886,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1336753,
    "ts": 1514656336186.761, "dur": 0.690,
    "args": {
      "External id": 2479887,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1336753,
    "ts": 1514656336189.353, "dur": 2.580,
    "args": {
      "External id": 2479888,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1336753,
    "ts": 1514656336193.205, "dur": 0.369,
    "args": {
      "External id": 2479889,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1336753,
    "ts": 1514656336195.675, "dur": 0.325,
    "args": {
      "External id": 2479890,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1336753,
    "ts": 1514656336199.802, "dur": 0.184,
    "args": {
      "External id": 2479891,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1336753,
    "ts": 1514656336215.781, "dur": 25.441,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656336301.734, "dur": 376.452,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656336335.653, "dur": 336.789,
    "args": {
      "External id": 2479894,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1336753,
    "ts": 1514656336346.057, "dur": 319.942,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656336706.670, "dur": 2.010,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.20)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656336799.108, "dur": 17168.818,
    "args": {
      "External id": 2479897,"Record function id": 0, "Ev Idx": 7600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656336896.048, "dur": 6.706,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656336906.017, "dur": 1.486,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656336909.247, "dur": 2.900,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656336913.891, "dur": 0.755,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656336916.040, "dur": 0.786,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656336918.253, "dur": 0.948,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656336923.085, "dur": 0.975,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656336925.747, "dur": 2.227,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656336929.726, "dur": 0.775,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656336932.072, "dur": 0.683,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656336953.301, "dur": 16968.742,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656336969.232, "dur": 16944.213,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656337042.455, "dur": 16.202,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656337061.808, "dur": 16816.351,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656337064.317, "dur": 16813.045,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656337070.779, "dur": 6.112,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656337078.751, "dur": 16795.066,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656354138.441, "dur": 54.745,
    "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": 1336753, "tid": 1336753, "ts": 1514656354138.441,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336753, "tid": 1336753,
    "ts": 1514656354175.997, "dur": 11.803,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656354181.807, "dur": 5.599,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656354260.075, "dur": 77.910,
    "args": {
      "External id": 2479918,"Record function id": 0, "Ev Idx": 7621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336753, "tid": 1336753,
    "ts": 1514656354339.492, "dur": 1104.164,
    "args": {
      "External id": 2479919,"Record function id": 0, "Ev Idx": 7622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656354381.891, "dur": 1047.988,
    "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": 1336753, "tid": 1336753, "ts": 1514656354381.891,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336753, "tid": 1336753,
    "ts": 1514656354447.346, "dur": 45.679,
    "args": {
      "External id": 2479921,"kernel_hash": "cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/ml/cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc.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": 1336753, "tid": 1336753,
    "ts": 1514656354506.080, "dur": 119.236,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656354640.260, "dur": 39.304,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656354687.214, "dur": 31.893,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656354749.285, "dur": 27.632,
    "args": {
      "External id": 2479925,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656354793.736, "dur": 16.776,
    "args": {
      "External id": 2479926,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656354830.204, "dur": 135.408,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656354880.232, "dur": 12.033,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656354886.010, "dur": 5.536,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656354895.231, "dur": 6.405,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656354903.143, "dur": 1.348,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656354909.521, "dur": 4.915,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656354976.507, "dur": 89.321,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656355103.024, "dur": 30.503,
    "args": {
      "External id": 2479934,"kernel_hash": "cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/w7/cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5.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": 1336753, "tid": 1336753,
    "ts": 1514656355144.616, "dur": 43.524,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656355194.457, "dur": 33.967,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656355253.289, "dur": 29.604,
    "args": {
      "External id": 2479937,"kernel_hash": "czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/zh/czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha.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": 1336753, "tid": 1336753,
    "ts": 1514656355288.861, "dur": 34.636,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656355340.552, "dur": 18.870,
    "args": {
      "External id": 2479939,"kernel_hash": "cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/nz/cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4.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": 1336753, "tid": 1336753,
    "ts": 1514656355506.418, "dur": 92.489,
    "args": {
      "External id": 2479940,"Record function id": 0, "Ev Idx": 7643
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336753, "tid": 1336753,
    "ts": 1514656355677.483, "dur": 48.094,
    "args": {
      "External id": 2479941,"Record function id": 0, "Ev Idx": 7644
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.21)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656355734.934, "dur": 18237.767,
    "args": {
      "External id": 2479942,"Record function id": 0, "Ev Idx": 7645
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.21)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656355743.949, "dur": 987.905,
    "args": {
      "External id": 2479943,"Record function id": 0, "Ev Idx": 7646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514656355823.586, "dur": 8.293,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656355846.582, "dur": 40.721,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656355852.515, "dur": 2.196,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656355859.285, "dur": 0.210,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656355861.540, "dur": 0.519,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656355864.122, "dur": 0.409,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656355868.229, "dur": 0.510,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656355870.390, "dur": 0.482,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656355872.455, "dur": 4.155,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656355878.232, "dur": 0.304,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656355880.373, "dur": 0.377,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656355899.884, "dur": 41.533,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656355974.692, "dur": 165.049,
    "args": {
      "External id": 2479956,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656356021.960, "dur": 5.794,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656356033.552, "dur": 10.962,
    "args": {
      "External id": 2479958,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656356037.884, "dur": 6.206,
    "args": {
      "External id": 2479959,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1336753,
    "ts": 1514656356041.786, "dur": 0.833,
    "args": {
      "External id": 2479960,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656356052.753, "dur": 34.018,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656356055.905, "dur": 2.757,
    "args": {
      "External id": 2479962,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656356060.254, "dur": 0.360,
    "args": {
      "External id": 2479963,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1336753,
    "ts": 1514656356062.103, "dur": 0.472,
    "args": {
      "External id": 2479964,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1336753,
    "ts": 1514656356066.945, "dur": 1.910,
    "args": {
      "External id": 2479965,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1336753,
    "ts": 1514656356070.365, "dur": 0.609,
    "args": {
      "External id": 2479966,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1336753,
    "ts": 1514656356072.486, "dur": 0.373,
    "args": {
      "External id": 2479967,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1336753,
    "ts": 1514656356076.848, "dur": 0.377,
    "args": {
      "External id": 2479968,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1336753,
    "ts": 1514656356078.689, "dur": 0.151,
    "args": {
      "External id": 2479969,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1336753,
    "ts": 1514656356080.572, "dur": 2.308,
    "args": {
      "External id": 2479970,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1336753,
    "ts": 1514656356101.934, "dur": 29.812,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656356241.114, "dur": 378.437,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656356283.300, "dur": 330.743,
    "args": {
      "External id": 2479973,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1336753,
    "ts": 1514656356297.168, "dur": 310.455,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656356648.639, "dur": 2.486,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.21)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656356752.462, "dur": 17015.241,
    "args": {
      "External id": 2479976,"Record function id": 0, "Ev Idx": 7679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656356856.735, "dur": 7.147,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656356867.342, "dur": 1.162,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656356869.998, "dur": 2.132,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656356873.827, "dur": 1.153,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656356876.553, "dur": 0.715,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656356878.712, "dur": 0.802,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656356883.054, "dur": 0.758,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656356885.440, "dur": 2.202,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656356889.113, "dur": 0.568,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656356891.411, "dur": 0.863,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656356912.526, "dur": 16810.856,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656356929.620, "dur": 16785.578,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656356951.622, "dur": 14.291,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656356968.835, "dur": 16711.368,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656356971.252, "dur": 16708.148,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656356977.531, "dur": 41.276,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656357021.615, "dur": 16654.915,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656373909.808, "dur": 37.044,
    "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": 1336753, "tid": 1336753, "ts": 1514656373909.808,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336753, "tid": 1336753,
    "ts": 1514656373933.172, "dur": 8.313,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656373936.527, "dur": 4.717,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656374042.774, "dur": 75.859,
    "args": {
      "External id": 2479997,"Record function id": 0, "Ev Idx": 7700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336753, "tid": 1336753,
    "ts": 1514656374120.756, "dur": 1108.114,
    "args": {
      "External id": 2479998,"Record function id": 0, "Ev Idx": 7701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656374164.609, "dur": 1051.403,
    "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": 1336753, "tid": 1336753, "ts": 1514656374164.609,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336753, "tid": 1336753,
    "ts": 1514656374235.004, "dur": 47.512,
    "args": {
      "External id": 2480000,"kernel_hash": "cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/ml/cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc.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": 1336753, "tid": 1336753,
    "ts": 1514656374295.866, "dur": 105.458,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656374411.234, "dur": 38.329,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656374457.798, "dur": 30.746,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656374514.731, "dur": 43.681,
    "args": {
      "External id": 2480004,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656374579.713, "dur": 17.539,
    "args": {
      "External id": 2480005,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656374616.547, "dur": 131.938,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656374665.879, "dur": 12.213,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656374671.188, "dur": 6.155,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656374681.008, "dur": 5.828,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656374688.212, "dur": 1.336,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656374692.036, "dur": 4.443,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656374758.758, "dur": 49.481,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656374840.127, "dur": 28.961,
    "args": {
      "External id": 2480013,"kernel_hash": "cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/w7/cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5.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": 1336753, "tid": 1336753,
    "ts": 1514656374878.023, "dur": 41.460,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656374928.176, "dur": 35.584,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656375022.571, "dur": 30.676,
    "args": {
      "External id": 2480016,"kernel_hash": "czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/zh/czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha.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": 1336753, "tid": 1336753,
    "ts": 1514656375060.188, "dur": 39.588,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656375121.414, "dur": 19.696,
    "args": {
      "External id": 2480018,"kernel_hash": "cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/nz/cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4.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": 1336753, "tid": 1336753,
    "ts": 1514656375292.234, "dur": 73.892,
    "args": {
      "External id": 2480019,"Record function id": 0, "Ev Idx": 7722
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336753, "tid": 1336753,
    "ts": 1514656375441.759, "dur": 46.357,
    "args": {
      "External id": 2480020,"Record function id": 0, "Ev Idx": 7723
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.22)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656375497.582, "dur": 18178.942,
    "args": {
      "External id": 2480021,"Record function id": 0, "Ev Idx": 7724
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.22)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656375505.333, "dur": 905.911,
    "args": {
      "External id": 2480022,"Record function id": 0, "Ev Idx": 7725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514656375601.313, "dur": 8.853,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656375625.190, "dur": 40.165,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656375631.052, "dur": 2.156,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656375638.494, "dur": 0.483,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656375640.566, "dur": 0.267,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656375642.604, "dur": 0.326,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656375646.517, "dur": 0.674,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656375649.023, "dur": 0.547,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656375651.020, "dur": 3.896,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656375656.433, "dur": 0.300,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656375657.922, "dur": 0.388,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656375677.705, "dur": 40.259,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656375752.701, "dur": 121.497,
    "args": {
      "External id": 2480035,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656375763.736, "dur": 4.429,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656375773.380, "dur": 10.490,
    "args": {
      "External id": 2480037,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656375777.906, "dur": 5.583,
    "args": {
      "External id": 2480038,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1336753,
    "ts": 1514656375781.790, "dur": 0.471,
    "args": {
      "External id": 2480039,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656375791.054, "dur": 39.512,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656375793.960, "dur": 2.685,
    "args": {
      "External id": 2480041,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656375798.246, "dur": 0.475,
    "args": {
      "External id": 2480042,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1336753,
    "ts": 1514656375800.305, "dur": 0.200,
    "args": {
      "External id": 2480043,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1336753,
    "ts": 1514656375805.019, "dur": 2.111,
    "args": {
      "External id": 2480044,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1336753,
    "ts": 1514656375808.647, "dur": 0.365,
    "args": {
      "External id": 2480045,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1336753,
    "ts": 1514656375811.272, "dur": 0.467,
    "args": {
      "External id": 2480046,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1336753,
    "ts": 1514656375815.345, "dur": 0.377,
    "args": {
      "External id": 2480047,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1336753,
    "ts": 1514656375817.608, "dur": 0.284,
    "args": {
      "External id": 2480048,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1336753,
    "ts": 1514656375823.169, "dur": 2.229,
    "args": {
      "External id": 2480049,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1336753,
    "ts": 1514656375841.472, "dur": 24.891,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656375926.377, "dur": 382.256,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656375960.272, "dur": 342.951,
    "args": {
      "External id": 2480052,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1336753,
    "ts": 1514656375970.712, "dur": 326.668,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656376336.709, "dur": 2.474,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.22)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656376432.680, "dur": 17027.691,
    "args": {
      "External id": 2480055,"Record function id": 0, "Ev Idx": 7758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656376552.889, "dur": 6.745,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656376563.491, "dur": 1.425,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656376566.677, "dur": 2.902,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656376571.820, "dur": 0.748,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656376574.043, "dur": 0.829,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656376576.028, "dur": 1.007,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656376581.038, "dur": 1.203,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656376583.735, "dur": 2.355,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656376587.857, "dur": 1.121,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656376590.931, "dur": 0.694,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656376613.499, "dur": 16807.380,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656376636.661, "dur": 16777.525,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656376657.162, "dur": 14.771,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656376674.758, "dur": 16705.435,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656376677.443, "dur": 16702.184,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656376683.508, "dur": 6.468,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656376691.522, "dur": 16685.352,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656393613.249, "dur": 36.976,
    "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": 1336753, "tid": 1336753, "ts": 1514656393613.249,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336753, "tid": 1336753,
    "ts": 1514656393636.626, "dur": 8.318,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656393639.930, "dur": 4.588,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656393716.625, "dur": 78.698,
    "args": {
      "External id": 2480076,"Record function id": 0, "Ev Idx": 7779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336753, "tid": 1336753,
    "ts": 1514656393796.839, "dur": 1085.677,
    "args": {
      "External id": 2480077,"Record function id": 0, "Ev Idx": 7780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656393839.521, "dur": 1029.521,
    "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": 1336753, "tid": 1336753, "ts": 1514656393839.521,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336753, "tid": 1336753,
    "ts": 1514656393905.102, "dur": 42.490,
    "args": {
      "External id": 2480079,"kernel_hash": "cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/ml/cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc.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": 1336753, "tid": 1336753,
    "ts": 1514656393960.965, "dur": 129.549,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656394104.390, "dur": 40.016,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656394153.079, "dur": 31.122,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656394211.472, "dur": 27.490,
    "args": {
      "External id": 2480083,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656394258.201, "dur": 17.428,
    "args": {
      "External id": 2480084,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656394295.121, "dur": 128.772,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656394343.902, "dur": 11.786,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656394349.482, "dur": 5.496,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656394358.813, "dur": 5.530,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656394365.773, "dur": 1.217,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656394369.249, "dur": 4.092,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656394434.373, "dur": 44.882,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656394511.375, "dur": 45.665,
    "args": {
      "External id": 2480092,"kernel_hash": "cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/w7/cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5.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": 1336753, "tid": 1336753,
    "ts": 1514656394569.913, "dur": 45.233,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656394624.104, "dur": 34.534,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656394683.967, "dur": 28.349,
    "args": {
      "External id": 2480095,"kernel_hash": "czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/zh/czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha.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": 1336753, "tid": 1336753,
    "ts": 1514656394718.410, "dur": 34.403,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656394774.079, "dur": 19.514,
    "args": {
      "External id": 2480097,"kernel_hash": "cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/nz/cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4.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": 1336753, "tid": 1336753,
    "ts": 1514656394945.346, "dur": 110.868,
    "args": {
      "External id": 2480098,"Record function id": 0, "Ev Idx": 7801
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336753, "tid": 1336753,
    "ts": 1514656395136.686, "dur": 49.140,
    "args": {
      "External id": 2480099,"Record function id": 0, "Ev Idx": 7802
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.23)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656395195.080, "dur": 18295.918,
    "args": {
      "External id": 2480100,"Record function id": 0, "Ev Idx": 7803
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.23)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656395204.004, "dur": 1000.368,
    "args": {
      "External id": 2480101,"Record function id": 0, "Ev Idx": 7804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514656395284.680, "dur": 8.617,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656395308.016, "dur": 41.112,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656395314.207, "dur": 2.347,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656395321.140, "dur": 0.271,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656395323.143, "dur": 0.412,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656395325.600, "dur": 0.436,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656395329.705, "dur": 0.538,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656395332.005, "dur": 0.410,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656395333.795, "dur": 4.405,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656395340.368, "dur": 0.315,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656395342.021, "dur": 0.555,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656395360.907, "dur": 44.131,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656395439.855, "dur": 219.809,
    "args": {
      "External id": 2480114,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656395450.546, "dur": 4.305,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656395459.882, "dur": 10.798,
    "args": {
      "External id": 2480116,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656395464.483, "dur": 5.791,
    "args": {
      "External id": 2480117,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1336753,
    "ts": 1514656395468.635, "dur": 0.533,
    "args": {
      "External id": 2480118,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656395477.295, "dur": 33.444,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656395480.088, "dur": 2.520,
    "args": {
      "External id": 2480120,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656395484.143, "dur": 0.201,
    "args": {
      "External id": 2480121,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1336753,
    "ts": 1514656395486.221, "dur": 0.200,
    "args": {
      "External id": 2480122,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1336753,
    "ts": 1514656395490.543, "dur": 1.623,
    "args": {
      "External id": 2480123,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1336753,
    "ts": 1514656395493.588, "dur": 0.322,
    "args": {
      "External id": 2480124,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1336753,
    "ts": 1514656395495.394, "dur": 0.503,
    "args": {
      "External id": 2480125,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1336753,
    "ts": 1514656395499.480, "dur": 0.735,
    "args": {
      "External id": 2480126,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1336753,
    "ts": 1514656395502.063, "dur": 0.150,
    "args": {
      "External id": 2480127,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1336753,
    "ts": 1514656395503.440, "dur": 2.514,
    "args": {
      "External id": 2480128,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1336753,
    "ts": 1514656395616.227, "dur": 32.939,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656395717.107, "dur": 387.766,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656395752.815, "dur": 346.555,
    "args": {
      "External id": 2480131,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1336753,
    "ts": 1514656395764.514, "dur": 326.224,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656396131.260, "dur": 2.233,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.23)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656396224.959, "dur": 17073.274,
    "args": {
      "External id": 2480134,"Record function id": 0, "Ev Idx": 7837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656396325.743, "dur": 6.674,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656396335.849, "dur": 1.611,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656396339.133, "dur": 2.565,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656396343.329, "dur": 0.925,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656396346.091, "dur": 1.193,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656396348.673, "dur": 0.676,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656396353.205, "dur": 0.717,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656396355.518, "dur": 2.160,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656396359.248, "dur": 0.935,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656396361.815, "dur": 0.820,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656396383.642, "dur": 16877.607,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656396401.376, "dur": 16852.696,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656396439.891, "dur": 12.808,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656396455.595, "dur": 16764.104,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656396458.167, "dur": 16761.030,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656396464.773, "dur": 5.846,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656396472.456, "dur": 16744.019,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656413428.937, "dur": 37.587,
    "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": 1336753, "tid": 1336753, "ts": 1514656413428.937,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336753, "tid": 1336753,
    "ts": 1514656413453.533, "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": 1336753, "tid": 1336753,
    "ts": 1514656413456.852, "dur": 4.285,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656413542.522, "dur": 78.672,
    "args": {
      "External id": 2480155,"Record function id": 0, "Ev Idx": 7858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336753, "tid": 1336753,
    "ts": 1514656413623.136, "dur": 1101.983,
    "args": {
      "External id": 2480156,"Record function id": 0, "Ev Idx": 7859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656413670.102, "dur": 1041.314,
    "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": 1336753, "tid": 1336753, "ts": 1514656413670.102,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336753, "tid": 1336753,
    "ts": 1514656413737.529, "dur": 45.542,
    "args": {
      "External id": 2480158,"kernel_hash": "cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/ml/cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc.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": 1336753, "tid": 1336753,
    "ts": 1514656413795.416, "dur": 101.919,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656413907.874, "dur": 38.779,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656413955.750, "dur": 68.161,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656414057.669, "dur": 29.765,
    "args": {
      "External id": 2480162,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656414106.521, "dur": 15.944,
    "args": {
      "External id": 2480163,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656414142.439, "dur": 130.235,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656414192.029, "dur": 12.248,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656414197.699, "dur": 5.783,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656414207.053, "dur": 5.417,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656414213.590, "dur": 1.149,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656414216.891, "dur": 4.701,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656414283.210, "dur": 48.703,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656414362.525, "dur": 27.531,
    "args": {
      "External id": 2480171,"kernel_hash": "cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/w7/cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5.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": 1336753, "tid": 1336753,
    "ts": 1514656414398.813, "dur": 41.286,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656414449.286, "dur": 34.159,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656414507.351, "dur": 40.793,
    "args": {
      "External id": 2480174,"kernel_hash": "czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/zh/czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha.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": 1336753, "tid": 1336753,
    "ts": 1514656414555.466, "dur": 38.444,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656414614.835, "dur": 19.636,
    "args": {
      "External id": 2480176,"kernel_hash": "cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/nz/cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4.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": 1336753, "tid": 1336753,
    "ts": 1514656414786.290, "dur": 75.419,
    "args": {
      "External id": 2480177,"Record function id": 0, "Ev Idx": 7880
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336753, "tid": 1336753,
    "ts": 1514656414936.264, "dur": 82.398,
    "args": {
      "External id": 2480178,"Record function id": 0, "Ev Idx": 7881
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.24)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656415030.983, "dur": 18273.503,
    "args": {
      "External id": 2480179,"Record function id": 0, "Ev Idx": 7882
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.24)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656415040.098, "dur": 909.103,
    "args": {
      "External id": 2480180,"Record function id": 0, "Ev Idx": 7883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514656415123.397, "dur": 8.954,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656415147.473, "dur": 35.970,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656415152.866, "dur": 2.257,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656415159.584, "dur": 0.275,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656415160.983, "dur": 0.426,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656415162.439, "dur": 0.395,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656415166.259, "dur": 0.536,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656415168.015, "dur": 0.511,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656415169.864, "dur": 3.802,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656415174.724, "dur": 0.341,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656415175.949, "dur": 0.291,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656415195.780, "dur": 40.822,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656415270.246, "dur": 114.435,
    "args": {
      "External id": 2480193,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656415280.956, "dur": 4.860,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656415291.156, "dur": 9.457,
    "args": {
      "External id": 2480195,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656415295.359, "dur": 4.863,
    "args": {
      "External id": 2480196,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1336753,
    "ts": 1514656415298.815, "dur": 0.319,
    "args": {
      "External id": 2480197,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656415307.512, "dur": 32.700,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656415310.040, "dur": 2.368,
    "args": {
      "External id": 2480199,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656415313.776, "dur": 0.412,
    "args": {
      "External id": 2480200,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1336753,
    "ts": 1514656415315.594, "dur": 0.264,
    "args": {
      "External id": 2480201,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1336753,
    "ts": 1514656415319.512, "dur": 2.154,
    "args": {
      "External id": 2480202,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1336753,
    "ts": 1514656415323.108, "dur": 0.512,
    "args": {
      "External id": 2480203,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1336753,
    "ts": 1514656415324.475, "dur": 0.580,
    "args": {
      "External id": 2480204,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1336753,
    "ts": 1514656415328.457, "dur": 0.401,
    "args": {
      "External id": 2480205,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1336753,
    "ts": 1514656415329.831, "dur": 0.337,
    "args": {
      "External id": 2480206,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1336753,
    "ts": 1514656415331.934, "dur": 2.552,
    "args": {
      "External id": 2480207,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1336753,
    "ts": 1514656415352.178, "dur": 24.485,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656415435.857, "dur": 412.540,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656415475.400, "dur": 367.551,
    "args": {
      "External id": 2480210,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1336753,
    "ts": 1514656415485.655, "dur": 351.308,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656415874.302, "dur": 2.146,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.24)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656415969.150, "dur": 17131.908,
    "args": {
      "External id": 2480213,"Record function id": 0, "Ev Idx": 7916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656416127.592, "dur": 7.062,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656416143.449, "dur": 1.133,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656416146.278, "dur": 2.499,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656416152.619, "dur": 0.562,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656416154.489, "dur": 0.762,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656416156.546, "dur": 0.696,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656416160.055, "dur": 0.710,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656416164.541, "dur": 3.179,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656416169.053, "dur": 1.087,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656416171.393, "dur": 0.676,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656416192.961, "dur": 16867.596,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656416208.360, "dur": 16845.509,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656416228.522, "dur": 15.558,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656416249.224, "dur": 16771.278,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656416251.635, "dur": 16768.343,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656416258.200, "dur": 5.704,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656416265.576, "dur": 16751.580,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656433241.965, "dur": 37.216,
    "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": 1336753, "tid": 1336753, "ts": 1514656433241.965,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336753, "tid": 1336753,
    "ts": 1514656433266.527, "dur": 7.577,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656433269.767, "dur": 4.122,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656433344.517, "dur": 75.353,
    "args": {
      "External id": 2480234,"Record function id": 0, "Ev Idx": 7937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336753, "tid": 1336753,
    "ts": 1514656433421.606, "dur": 1102.137,
    "args": {
      "External id": 2480235,"Record function id": 0, "Ev Idx": 7938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656433460.420, "dur": 1050.039,
    "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": 1336753, "tid": 1336753, "ts": 1514656433460.420,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336753, "tid": 1336753,
    "ts": 1514656433525.261, "dur": 58.021,
    "args": {
      "External id": 2480237,"kernel_hash": "cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/ml/cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc.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": 1336753, "tid": 1336753,
    "ts": 1514656433598.877, "dur": 104.724,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656433713.142, "dur": 37.636,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656433759.770, "dur": 30.702,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656433816.384, "dur": 25.595,
    "args": {
      "External id": 2480241,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656433860.649, "dur": 17.482,
    "args": {
      "External id": 2480242,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656433895.637, "dur": 167.205,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656433943.581, "dur": 11.384,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656433948.613, "dur": 5.507,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656433957.649, "dur": 5.611,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656433964.629, "dur": 1.382,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656433968.155, "dur": 3.682,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656434075.186, "dur": 51.266,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656434158.847, "dur": 30.045,
    "args": {
      "External id": 2480250,"kernel_hash": "cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/w7/cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5.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": 1336753, "tid": 1336753,
    "ts": 1514656434197.725, "dur": 42.156,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656434248.691, "dur": 34.688,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656434330.428, "dur": 29.860,
    "args": {
      "External id": 2480253,"kernel_hash": "czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/zh/czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha.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": 1336753, "tid": 1336753,
    "ts": 1514656434368.986, "dur": 35.378,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656434424.649, "dur": 18.231,
    "args": {
      "External id": 2480255,"kernel_hash": "cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/nz/cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4.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": 1336753, "tid": 1336753,
    "ts": 1514656434604.352, "dur": 73.663,
    "args": {
      "External id": 2480256,"Record function id": 0, "Ev Idx": 7959
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336753, "tid": 1336753,
    "ts": 1514656434753.299, "dur": 47.283,
    "args": {
      "External id": 2480257,"Record function id": 0, "Ev Idx": 7960
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.25)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656434808.908, "dur": 18271.892,
    "args": {
      "External id": 2480258,"Record function id": 0, "Ev Idx": 7961
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.25)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656434817.117, "dur": 922.923,
    "args": {
      "External id": 2480259,"Record function id": 0, "Ev Idx": 7962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514656434896.811, "dur": 9.021,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656434919.732, "dur": 35.066,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656434925.418, "dur": 2.155,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656434932.216, "dur": 0.265,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656434933.715, "dur": 0.543,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656434935.042, "dur": 0.248,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656434938.191, "dur": 0.359,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656434939.417, "dur": 0.363,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656434940.649, "dur": 3.327,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656434945.103, "dur": 0.160,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656434946.731, "dur": 0.153,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656434966.727, "dur": 74.137,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656435079.526, "dur": 113.359,
    "args": {
      "External id": 2480272,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656435090.863, "dur": 5.901,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656435102.060, "dur": 10.020,
    "args": {
      "External id": 2480274,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656435106.315, "dur": 5.395,
    "args": {
      "External id": 2480275,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1336753,
    "ts": 1514656435109.578, "dur": 0.623,
    "args": {
      "External id": 2480276,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656435119.690, "dur": 27.278,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656435121.826, "dur": 2.287,
    "args": {
      "External id": 2480278,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656435125.483, "dur": 0.306,
    "args": {
      "External id": 2480279,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1336753,
    "ts": 1514656435127.288, "dur": 0.322,
    "args": {
      "External id": 2480280,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1336753,
    "ts": 1514656435130.915, "dur": 1.181,
    "args": {
      "External id": 2480281,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1336753,
    "ts": 1514656435133.035, "dur": 0.275,
    "args": {
      "External id": 2480282,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1336753,
    "ts": 1514656435134.548, "dur": 0.307,
    "args": {
      "External id": 2480283,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1336753,
    "ts": 1514656435137.853, "dur": 0.276,
    "args": {
      "External id": 2480284,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1336753,
    "ts": 1514656435138.954, "dur": 0.163,
    "args": {
      "External id": 2480285,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1336753,
    "ts": 1514656435139.856, "dur": 2.459,
    "args": {
      "External id": 2480286,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1336753,
    "ts": 1514656435159.970, "dur": 25.406,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656435245.359, "dur": 395.209,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656435278.998, "dur": 356.041,
    "args": {
      "External id": 2480289,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1336753,
    "ts": 1514656435289.021, "dur": 339.519,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656435666.182, "dur": 2.107,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.25)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656435761.199, "dur": 17092.580,
    "args": {
      "External id": 2480292,"Record function id": 0, "Ev Idx": 7995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656435864.777, "dur": 6.601,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656435874.916, "dur": 1.174,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656435877.779, "dur": 2.499,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656435881.959, "dur": 0.684,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656435883.986, "dur": 0.834,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656435886.212, "dur": 0.713,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656435890.506, "dur": 1.002,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656435892.631, "dur": 2.484,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656435896.469, "dur": 0.756,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656435899.050, "dur": 0.747,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656435920.494, "dur": 16894.907,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656435936.845, "dur": 16871.665,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656435962.181, "dur": 14.980,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656435979.821, "dur": 16796.017,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656436020.819, "dur": 16754.167,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656436027.305, "dur": 6.282,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656436035.333, "dur": 16737.038,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656453013.571, "dur": 37.455,
    "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": 1336753, "tid": 1336753, "ts": 1514656453013.571,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336753, "tid": 1336753,
    "ts": 1514656453037.331, "dur": 8.314,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656453040.613, "dur": 4.645,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656453119.906, "dur": 77.656,
    "args": {
      "External id": 2480313,"Record function id": 0, "Ev Idx": 8016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336753, "tid": 1336753,
    "ts": 1514656453199.160, "dur": 1113.074,
    "args": {
      "External id": 2480314,"Record function id": 0, "Ev Idx": 8017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656453241.721, "dur": 1056.714,
    "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": 1336753, "tid": 1336753, "ts": 1514656453241.721,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336753, "tid": 1336753,
    "ts": 1514656453312.513, "dur": 45.586,
    "args": {
      "External id": 2480316,"kernel_hash": "cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/ml/cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc.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": 1336753, "tid": 1336753,
    "ts": 1514656453370.658, "dur": 103.005,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656453483.529, "dur": 41.790,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656453551.520, "dur": 35.543,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656453616.661, "dur": 29.907,
    "args": {
      "External id": 2480320,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656453664.947, "dur": 15.933,
    "args": {
      "External id": 2480321,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656453699.320, "dur": 127.841,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656453749.503, "dur": 10.859,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656453754.262, "dur": 5.236,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656453763.442, "dur": 4.593,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656453769.191, "dur": 1.218,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656453772.769, "dur": 3.997,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656453836.697, "dur": 45.006,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656453911.103, "dur": 30.164,
    "args": {
      "External id": 2480329,"kernel_hash": "cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/w7/cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5.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": 1336753, "tid": 1336753,
    "ts": 1514656453949.880, "dur": 77.214,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656454039.247, "dur": 40.893,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656454105.280, "dur": 40.589,
    "args": {
      "External id": 2480332,"kernel_hash": "czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/zh/czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha.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": 1336753, "tid": 1336753,
    "ts": 1514656454151.593, "dur": 34.551,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656454206.795, "dur": 20.231,
    "args": {
      "External id": 2480334,"kernel_hash": "cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/nz/cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4.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": 1336753, "tid": 1336753,
    "ts": 1514656454375.438, "dur": 72.744,
    "args": {
      "External id": 2480335,"Record function id": 0, "Ev Idx": 8038
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336753, "tid": 1336753,
    "ts": 1514656454523.000, "dur": 67.288,
    "args": {
      "External id": 2480336,"Record function id": 0, "Ev Idx": 8039
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.26)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656454600.501, "dur": 18223.345,
    "args": {
      "External id": 2480337,"Record function id": 0, "Ev Idx": 8040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.26)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656454611.089, "dur": 951.037,
    "args": {
      "External id": 2480338,"Record function id": 0, "Ev Idx": 8041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514656454691.353, "dur": 9.064,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656454715.012, "dur": 33.812,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656454720.751, "dur": 2.184,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656454727.158, "dur": 0.801,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656454728.800, "dur": 0.530,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656454730.320, "dur": 0.461,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656454733.695, "dur": 0.737,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656454735.388, "dur": 0.603,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656454736.841, "dur": 2.441,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656454740.056, "dur": 0.551,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656454741.879, "dur": 0.481,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656454760.306, "dur": 43.975,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656454837.321, "dur": 128.341,
    "args": {
      "External id": 2480351,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656454847.740, "dur": 4.586,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656454857.768, "dur": 10.700,
    "args": {
      "External id": 2480353,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656454862.233, "dur": 5.817,
    "args": {
      "External id": 2480354,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1336753,
    "ts": 1514656454866.166, "dur": 0.748,
    "args": {
      "External id": 2480355,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656454875.424, "dur": 47.074,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656454877.774, "dur": 3.066,
    "args": {
      "External id": 2480357,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656454882.241, "dur": 0.649,
    "args": {
      "External id": 2480358,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1336753,
    "ts": 1514656454899.866, "dur": 1.467,
    "args": {
      "External id": 2480359,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1336753,
    "ts": 1514656454904.804, "dur": 0.798,
    "args": {
      "External id": 2480360,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1336753,
    "ts": 1514656454906.792, "dur": 0.679,
    "args": {
      "External id": 2480361,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1336753,
    "ts": 1514656454908.748, "dur": 2.697,
    "args": {
      "External id": 2480362,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1336753,
    "ts": 1514656454912.606, "dur": 0.501,
    "args": {
      "External id": 2480363,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1336753,
    "ts": 1514656454913.708, "dur": 0.494,
    "args": {
      "External id": 2480364,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1336753,
    "ts": 1514656454917.257, "dur": 0.598,
    "args": {
      "External id": 2480365,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1336753,
    "ts": 1514656454932.994, "dur": 25.036,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656455053.840, "dur": 392.853,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656455089.208, "dur": 352.572,
    "args": {
      "External id": 2480368,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1336753,
    "ts": 1514656455099.367, "dur": 335.781,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656455471.825, "dur": 2.465,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.26)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656455588.282, "dur": 17025.733,
    "args": {
      "External id": 2480371,"Record function id": 0, "Ev Idx": 8074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656455689.808, "dur": 6.913,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656455699.802, "dur": 1.156,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656455709.200, "dur": 1.246,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656455714.099, "dur": 1.049,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656455716.573, "dur": 1.376,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656455719.194, "dur": 1.012,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656455721.775, "dur": 1.007,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656455726.455, "dur": 2.145,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656455729.805, "dur": 1.511,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656455732.860, "dur": 0.996,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656455753.142, "dur": 16817.240,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656455768.915, "dur": 16793.771,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656455786.853, "dur": 13.647,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656455805.467, "dur": 16709.767,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656455808.084, "dur": 16706.381,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656455814.221, "dur": 5.850,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656455821.634, "dur": 16689.868,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656472759.574, "dur": 39.368,
    "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": 1336753, "tid": 1336753, "ts": 1514656472759.574,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336753, "tid": 1336753,
    "ts": 1514656472784.008, "dur": 9.706,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656472787.897, "dur": 5.613,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656472864.381, "dur": 78.713,
    "args": {
      "External id": 2480392,"Record function id": 0, "Ev Idx": 8095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336753, "tid": 1336753,
    "ts": 1514656472944.794, "dur": 1136.500,
    "args": {
      "External id": 2480393,"Record function id": 0, "Ev Idx": 8096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656473012.272, "dur": 1054.491,
    "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": 1336753, "tid": 1336753, "ts": 1514656473012.272,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336753, "tid": 1336753,
    "ts": 1514656473087.127, "dur": 48.032,
    "args": {
      "External id": 2480395,"kernel_hash": "cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/ml/cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc.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": 1336753, "tid": 1336753,
    "ts": 1514656473149.664, "dur": 103.213,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656473261.568, "dur": 38.500,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656473309.143, "dur": 32.093,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656473366.546, "dur": 28.334,
    "args": {
      "External id": 2480399,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656473416.636, "dur": 18.034,
    "args": {
      "External id": 2480400,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656473452.491, "dur": 152.571,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656473503.011, "dur": 12.402,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656473508.055, "dur": 6.471,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656473518.231, "dur": 5.078,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656473524.466, "dur": 1.715,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656473544.826, "dur": 4.875,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656473617.594, "dur": 51.655,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656473702.338, "dur": 30.875,
    "args": {
      "External id": 2480408,"kernel_hash": "cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/w7/cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5.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": 1336753, "tid": 1336753,
    "ts": 1514656473742.521, "dur": 42.244,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656473793.505, "dur": 33.854,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656473848.996, "dur": 28.239,
    "args": {
      "External id": 2480411,"kernel_hash": "czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/zh/czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha.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": 1336753, "tid": 1336753,
    "ts": 1514656473882.643, "dur": 33.600,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656473937.687, "dur": 18.872,
    "args": {
      "External id": 2480413,"kernel_hash": "cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/nz/cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4.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": 1336753, "tid": 1336753,
    "ts": 1514656474148.122, "dur": 72.767,
    "args": {
      "External id": 2480414,"Record function id": 0, "Ev Idx": 8117
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336753, "tid": 1336753,
    "ts": 1514656474297.133, "dur": 47.939,
    "args": {
      "External id": 2480415,"Record function id": 0, "Ev Idx": 8118
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.27)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656474353.871, "dur": 18202.218,
    "args": {
      "External id": 2480416,"Record function id": 0, "Ev Idx": 8119
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.27)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656474362.652, "dur": 913.564,
    "args": {
      "External id": 2480417,"Record function id": 0, "Ev Idx": 8120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514656474443.941, "dur": 9.511,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656474468.402, "dur": 33.989,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656474473.560, "dur": 2.305,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656474480.709, "dur": 0.473,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656474482.459, "dur": 0.788,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656474484.166, "dur": 0.527,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656474487.600, "dur": 0.734,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656474489.407, "dur": 0.600,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656474490.859, "dur": 2.464,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656474494.360, "dur": 0.676,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656474495.809, "dur": 0.436,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656474513.712, "dur": 57.055,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656474608.005, "dur": 113.405,
    "args": {
      "External id": 2480430,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656474619.162, "dur": 5.195,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656474629.552, "dur": 10.907,
    "args": {
      "External id": 2480432,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "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": 1336753, "tid": 1336753,
    "ts": 1514656474634.189, "dur": 5.849,
    "args": {
      "External id": 2480433,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "6423040", "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": 1336753, "tid": 1336753,
    "ts": 1514656474637.404, "dur": 1.059,
    "args": {
      "External id": 2480434,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656474647.459, "dur": 26.838,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656474649.402, "dur": 2.806,
    "args": {
      "External id": 2480436,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "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": 1336753, "tid": 1336753,
    "ts": 1514656474653.171, "dur": 0.613,
    "args": {
      "External id": 2480437,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "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": 1336753, "tid": 1336753,
    "ts": 1514656474654.606, "dur": 0.444,
    "args": {
      "External id": 2480438,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "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": 1336753, "tid": 1336753,
    "ts": 1514656474658.319, "dur": 0.569,
    "args": {
      "External id": 2480439,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "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": 1336753, "tid": 1336753,
    "ts": 1514656474659.733, "dur": 0.559,
    "args": {
      "External id": 2480440,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "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": 1336753, "tid": 1336753,
    "ts": 1514656474661.344, "dur": 0.428,
    "args": {
      "External id": 2480441,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "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": 1336753, "tid": 1336753,
    "ts": 1514656474664.488, "dur": 0.411,
    "args": {
      "External id": 2480442,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "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": 1336753, "tid": 1336753,
    "ts": 1514656474665.728, "dur": 0.590,
    "args": {
      "External id": 2480443,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "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": 1336753, "tid": 1336753,
    "ts": 1514656474667.410, "dur": 2.765,
    "args": {
      "External id": 2480444,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "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": 1336753, "tid": 1336753,
    "ts": 1514656474687.181, "dur": 26.150,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656474773.271, "dur": 400.536,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656474806.224, "dur": 362.289,
    "args": {
      "External id": 2480447,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 0, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1336753,
    "ts": 1514656474816.477, "dur": 345.335,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656475200.318, "dur": 2.219,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.27)", "pid": 1336753, "tid": 1336753,
    "ts": 1514656475296.289, "dur": 17041.479,
    "args": {
      "External id": 2480450,"Record function id": 0, "Ev Idx": 8153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656475438.590, "dur": 8.429,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656475450.874, "dur": 1.484,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656475454.174, "dur": 0.961,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656475459.028, "dur": 1.670,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656475462.063, "dur": 1.255,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656475464.543, "dur": 1.488,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656475467.792, "dur": 1.251,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656475472.529, "dur": 2.627,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656475476.491, "dur": 1.030,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656475478.897, "dur": 1.200,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656475500.474, "dur": 16794.628,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656475516.010, "dur": 16772.332,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656475550.302, "dur": 14.560,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656475570.017, "dur": 16683.312,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656475572.459, "dur": 16680.275,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656475578.375, "dur": 6.203,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656475586.232, "dur": 16663.648,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656492479.969, "dur": 37.953,
    "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": 1336753, "tid": 1336753, "ts": 1514656492479.969,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336753, "tid": 1336753,
    "ts": 1514656492504.069, "dur": 8.640,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656492507.462, "dur": 5.005,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656492599.622, "dur": 76.591,
    "args": {
      "External id": 2480471,"Record function id": 0, "Ev Idx": 8174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336753, "tid": 1336753,
    "ts": 1514656492677.920, "dur": 1100.920,
    "args": {
      "External id": 2480472,"Record function id": 0, "Ev Idx": 8175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656492718.386, "dur": 1047.773,
    "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": 1336753, "tid": 1336753, "ts": 1514656492718.386,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336753, "tid": 1336753,
    "ts": 1514656492789.441, "dur": 45.951,
    "args": {
      "External id": 2480474,"kernel_hash": "cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/ml/cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc.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": 1336753, "tid": 1336753,
    "ts": 1514656492848.285, "dur": 103.310,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656492960.639, "dur": 75.491,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656493049.888, "dur": 34.275,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656493110.786, "dur": 29.298,
    "args": {
      "External id": 2480478,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656493160.634, "dur": 16.368,
    "args": {
      "External id": 2480479,"kernel_hash": "cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554", "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/ex/cex6azrfnirqh3jognpirazamj4oyxhnk3mxs7t2vmitx2fbo554.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": 1336753, "tid": 1336753,
    "ts": 1514656493195.221, "dur": 128.244,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656493245.412, "dur": 12.000,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656493250.702, "dur": 6.031,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656493260.165, "dur": 5.144,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656493266.828, "dur": 1.616,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656493270.847, "dur": 3.064,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656493334.630, "dur": 47.474,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656493415.370, "dur": 29.635,
    "args": {
      "External id": 2480487,"kernel_hash": "cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/w7/cw7dgbt4tjenpklhtvob3eghj7bbox7g2wxbo732slxqx2vrsyd5.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": 1336753, "tid": 1336753,
    "ts": 1514656493453.793, "dur": 40.598,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656493503.372, "dur": 51.280,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656493581.144, "dur": 29.502,
    "args": {
      "External id": 2480490,"kernel_hash": "czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/zh/czh4pfojeoszm3aid3si2yxrotnl4lfradxhc2mhjrcx6kyggzha.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": 1336753, "tid": 1336753,
    "ts": 1514656493616.940, "dur": 38.035,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656493677.157, "dur": 18.937,
    "args": {
      "External id": 2480492,"kernel_hash": "cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/nz/cnzai7gm7fkymyw2bih55vxu3i63al2yxbe4zf3zad5ca4tx2gz4.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": 1336753, "tid": 1336753,
    "ts": 1514656493841.396, "dur": 32.783,
    "args": {
      "External id": 2480493,"Record function id": 0, "Ev Idx": 8196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656494051.624, "dur": 300.622,
    "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": 1336753, "tid": 1336753, "ts": 1514656494051.624,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1336753,
    "ts": 1514656494083.401, "dur": 8.583,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656494085.437, "dur": 6.077,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656494103.275, "dur": 12.645,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656494106.577, "dur": 8.727,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656494126.182, "dur": 5.474,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656494331.558, "dur": 6.894,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656494335.665, "dur": 2.500,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656494383.103, "dur": 128.936,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656494385.776, "dur": 15.578,
    "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": 1336753, "tid": 1336753, "ts": 1514656494385.776,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336753, "tid": 1336753,
    "ts": 1514656494391.388, "dur": 8.245,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656494396.996, "dur": 2.274,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656494403.795, "dur": 107.956,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656494406.415, "dur": 5.293,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656494407.569, "dur": 3.999,
    "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": 1336753, "tid": 1336753, "ts": 1514656494407.569,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1336753,
    "ts": 1514656494415.831, "dur": 86.990,
    "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": 1336753, "tid": 1336753, "ts": 1514656494415.831,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656494506.301, "dur": 4.376,
    "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": 1336753, "tid": 1336753, "ts": 1514656494506.301,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336753, "tid": 1336753,
    "ts": 1514656494521.743, "dur": 95.603,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656494522.589, "dur": 26.302,
    "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": 1336753, "tid": 1336753, "ts": 1514656494522.589,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336753, "tid": 1336753,
    "ts": 1514656494524.378, "dur": 22.694,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656494545.108, "dur": 1.450,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656494549.829, "dur": 67.311,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656494551.289, "dur": 6.140,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656494552.095, "dur": 5.167,
    "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": 1336753, "tid": 1336753, "ts": 1514656494552.095,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1336753,
    "ts": 1514656494558.019, "dur": 51.603,
    "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": 1336753, "tid": 1336753, "ts": 1514656494558.019,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656494611.279, "dur": 5.295,
    "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": 1336753, "tid": 1336753, "ts": 1514656494611.279,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336753, "tid": 1336753,
    "ts": 1514656494626.420, "dur": 65.681,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656494627.078, "dur": 5.872,
    "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": 1336753, "tid": 1336753, "ts": 1514656494627.078,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336753, "tid": 1336753,
    "ts": 1514656494629.058, "dur": 2.753,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656494630.783, "dur": 0.821,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656494636.186, "dur": 55.728,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656494636.975, "dur": 4.632,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656494637.793, "dur": 3.668,
    "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": 1336753, "tid": 1336753, "ts": 1514656494637.793,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1336753,
    "ts": 1514656494642.190, "dur": 41.368,
    "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": 1336753, "tid": 1336753, "ts": 1514656494642.190,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656494685.154, "dur": 6.381,
    "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": 1336753, "tid": 1336753, "ts": 1514656494685.154,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1336753,
    "ts": 1514656494717.100, "dur": 4.321,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656494718.410, "dur": 2.869,
    "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": 1336753, "tid": 1336753, "ts": 1514656494718.410,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1336753,
    "ts": 1514656494728.559, "dur": 3.991,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656494729.798, "dur": 2.627,
    "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": 1336753, "tid": 1336753, "ts": 1514656494729.798,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1336753,
    "ts": 1514656494739.515, "dur": 5.398,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656494740.847, "dur": 3.919,
    "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": 1336753, "tid": 1336753, "ts": 1514656494740.847,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656494781.472, "dur": 187.644,
    "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": 1336753, "tid": 1336753, "ts": 1514656494781.472,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336753, "tid": 1336753,
    "ts": 1514656494803.401, "dur": 9.992,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656494807.170, "dur": 5.602,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495015.772, "dur": 123.298,
    "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": 1336753, "tid": 1336753, "ts": 1514656495015.772,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336753, "tid": 1336753,
    "ts": 1514656495032.603, "dur": 9.430,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495035.486, "dur": 5.960,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495169.602, "dur": 193.315,
    "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": 1336753, "tid": 1336753, "ts": 1514656495169.602,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336753, "tid": 1336753,
    "ts": 1514656495200.079, "dur": 133.187,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495251.787, "dur": 7.435,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495254.316, "dur": 4.271,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495262.067, "dur": 4.563,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495267.944, "dur": 1.897,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495274.645, "dur": 3.799,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495347.174, "dur": 5.731,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495368.751, "dur": 6.659,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495370.554, "dur": 4.609,
    "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": 1336753, "tid": 1336753, "ts": 1514656495370.554,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336753, "tid": 1336753,
    "ts": 1514656495394.760, "dur": 124.471,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495398.467, "dur": 8.879,
    "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": 1336753, "tid": 1336753, "ts": 1514656495398.467,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336753, "tid": 1336753,
    "ts": 1514656495401.483, "dur": 4.859,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495403.857, "dur": 2.193,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495408.666, "dur": 110.193,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495410.869, "dur": 9.625,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495413.932, "dur": 6.400,
    "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": 1336753, "tid": 1336753, "ts": 1514656495413.932,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1336753,
    "ts": 1514656495421.485, "dur": 88.419,
    "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": 1336753, "tid": 1336753, "ts": 1514656495421.485,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656495512.401, "dur": 5.717,
    "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": 1336753, "tid": 1336753, "ts": 1514656495512.401,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656495574.978, "dur": 233.300,
    "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": 1336753, "tid": 1336753, "ts": 1514656495574.978,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1336753,
    "ts": 1514656495597.968, "dur": 3.773,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495598.983, "dur": 2.465,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495607.476, "dur": 3.213,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495608.616, "dur": 1.956,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495609.381, "dur": 1.097,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495618.753, "dur": 10.156,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495623.149, "dur": 5.402,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495635.554, "dur": 3.567,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495642.597, "dur": 3.401,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495782.251, "dur": 3.846,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495783.656, "dur": 2.205,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495788.750, "dur": 2.774,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495789.969, "dur": 1.438,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495830.642, "dur": 102.061,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495831.804, "dur": 8.680,
    "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": 1336753, "tid": 1336753, "ts": 1514656495831.804,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336753, "tid": 1336753,
    "ts": 1514656495834.811, "dur": 4.337,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495836.839, "dur": 2.037,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495841.397, "dur": 91.009,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495845.444, "dur": 4.010,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495846.241, "dur": 3.083,
    "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": 1336753, "tid": 1336753, "ts": 1514656495846.241,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1336753,
    "ts": 1514656495850.221, "dur": 75.599,
    "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": 1336753, "tid": 1336753, "ts": 1514656495850.221,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656495928.499, "dur": 3.288,
    "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": 1336753, "tid": 1336753, "ts": 1514656495928.499,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336753, "tid": 1336753,
    "ts": 1514656495941.663, "dur": 121.797,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495944.941, "dur": 6.798,
    "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": 1336753, "tid": 1336753, "ts": 1514656495944.941,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336753, "tid": 1336753,
    "ts": 1514656495947.351, "dur": 3.066,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495949.153, "dur": 1.067,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495952.360, "dur": 110.882,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495953.478, "dur": 8.749,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656495956.692, "dur": 5.383,
    "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": 1336753, "tid": 1336753, "ts": 1514656495956.692,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1336753,
    "ts": 1514656495963.036, "dur": 89.707,
    "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": 1336753, "tid": 1336753, "ts": 1514656495963.036,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656496056.573, "dur": 6.037,
    "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": 1336753, "tid": 1336753, "ts": 1514656496056.573,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SwiGLULinearFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656496091.442, "dur": 173.736,
    "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": 1336753, "tid": 1336753, "ts": 1514656496091.442,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514656496140.955, "dur": 5.398,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496182.226, "dur": 67.459,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496183.538, "dur": 6.713,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496184.980, "dur": 4.303,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496187.368, "dur": 1.660,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496191.055, "dur": 58.289,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496194.789, "dur": 3.077,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496196.004, "dur": 1.749,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496198.521, "dur": 47.026,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496247.344, "dur": 1.181,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496274.983, "dur": 28.971,
    "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": 1336753, "tid": 1336753, "ts": 1514656496274.983,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656496339.243, "dur": 187.352,
    "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": 1336753, "tid": 1336753, "ts": 1514656496339.243,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1336753,
    "ts": 1514656496359.050, "dur": 4.941,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496362.100, "dur": 1.707,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496372.309, "dur": 7.520,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496374.743, "dur": 4.601,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496386.397, "dur": 3.603,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496511.651, "dur": 3.154,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496512.776, "dur": 1.828,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496564.748, "dur": 96.660,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496565.964, "dur": 11.809,
    "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": 1336753, "tid": 1336753, "ts": 1514656496565.964,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336753, "tid": 1336753,
    "ts": 1514656496571.580, "dur": 4.257,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496573.475, "dur": 2.093,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496578.941, "dur": 82.196,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496580.515, "dur": 5.843,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496583.279, "dur": 2.960,
    "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": 1336753, "tid": 1336753, "ts": 1514656496583.279,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1336753,
    "ts": 1514656496587.046, "dur": 68.207,
    "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": 1336753, "tid": 1336753, "ts": 1514656496587.046,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656496657.325, "dur": 3.118,
    "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": 1336753, "tid": 1336753, "ts": 1514656496657.325,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336753, "tid": 1336753,
    "ts": 1514656496670.842, "dur": 66.799,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496671.654, "dur": 10.517,
    "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": 1336753, "tid": 1336753, "ts": 1514656496671.654,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336753, "tid": 1336753,
    "ts": 1514656496678.317, "dur": 2.686,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496679.911, "dur": 0.779,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496682.874, "dur": 54.584,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496683.577, "dur": 5.273,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496684.543, "dur": 4.170,
    "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": 1336753, "tid": 1336753, "ts": 1514656496684.543,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1336753,
    "ts": 1514656496689.481, "dur": 43.721,
    "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": 1336753, "tid": 1336753, "ts": 1514656496689.481,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656496734.844, "dur": 2.278,
    "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": 1336753, "tid": 1336753, "ts": 1514656496734.844,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336753, "tid": 1336753,
    "ts": 1514656496744.480, "dur": 63.385,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496745.123, "dur": 9.356,
    "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": 1336753, "tid": 1336753, "ts": 1514656496745.123,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336753, "tid": 1336753,
    "ts": 1514656496746.635, "dur": 6.765,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496750.310, "dur": 2.879,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496755.395, "dur": 52.262,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496756.565, "dur": 4.355,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496757.193, "dur": 3.594,
    "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": 1336753, "tid": 1336753, "ts": 1514656496757.193,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1336753,
    "ts": 1514656496763.784, "dur": 37.678,
    "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": 1336753, "tid": 1336753, "ts": 1514656496763.784,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656496802.810, "dur": 4.485,
    "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": 1336753, "tid": 1336753, "ts": 1514656496802.810,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1336753,
    "ts": 1514656496824.460, "dur": 3.743,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496825.102, "dur": 2.945,
    "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": 1336753, "tid": 1336753, "ts": 1514656496825.102,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1336753,
    "ts": 1514656496910.688, "dur": 8.510,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496914.466, "dur": 4.578,
    "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": 1336753, "tid": 1336753, "ts": 1514656496914.466,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1336753,
    "ts": 1514656496924.812, "dur": 4.481,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496925.521, "dur": 3.598,
    "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": 1336753, "tid": 1336753, "ts": 1514656496925.521,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656496959.012, "dur": 201.863,
    "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": 1336753, "tid": 1336753, "ts": 1514656496959.012,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336753, "tid": 1336753,
    "ts": 1514656496978.067, "dur": 41.809,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656496980.505, "dur": 38.466,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497176.818, "dur": 109.262,
    "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": 1336753, "tid": 1336753, "ts": 1514656497176.818,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336753, "tid": 1336753,
    "ts": 1514656497192.537, "dur": 8.074,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497195.223, "dur": 5.048,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497312.957, "dur": 178.135,
    "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": 1336753, "tid": 1336753, "ts": 1514656497312.957,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336753, "tid": 1336753,
    "ts": 1514656497341.320, "dur": 122.938,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497389.917, "dur": 7.168,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497392.308, "dur": 4.380,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497399.536, "dur": 4.485,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497405.369, "dur": 3.469,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497410.994, "dur": 3.359,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497476.038, "dur": 5.106,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497496.808, "dur": 8.528,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497498.546, "dur": 6.604,
    "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": 1336753, "tid": 1336753, "ts": 1514656497498.546,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336753, "tid": 1336753,
    "ts": 1514656497522.327, "dur": 141.621,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497523.730, "dur": 29.138,
    "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": 1336753, "tid": 1336753, "ts": 1514656497523.730,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336753, "tid": 1336753,
    "ts": 1514656497526.235, "dur": 24.972,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497548.614, "dur": 2.017,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497554.240, "dur": 109.342,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497558.569, "dur": 6.948,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497559.407, "dur": 5.903,
    "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": 1336753, "tid": 1336753, "ts": 1514656497559.407,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1336753,
    "ts": 1514656497566.421, "dur": 87.565,
    "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": 1336753, "tid": 1336753, "ts": 1514656497566.421,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656497656.497, "dur": 6.309,
    "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": 1336753, "tid": 1336753, "ts": 1514656497656.497,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656497701.843, "dur": 214.251,
    "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": 1336753, "tid": 1336753, "ts": 1514656497701.843,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1336753,
    "ts": 1514656497723.293, "dur": 3.053,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497724.246, "dur": 1.786,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497730.273, "dur": 5.122,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497731.182, "dur": 4.110,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497734.167, "dur": 0.973,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497743.094, "dur": 7.315,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497745.358, "dur": 4.746,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497756.666, "dur": 3.209,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497763.144, "dur": 3.687,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497891.203, "dur": 3.281,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497892.404, "dur": 1.862,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497899.411, "dur": 2.482,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497900.455, "dur": 1.315,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497936.666, "dur": 138.311,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497937.915, "dur": 8.368,
    "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": 1336753, "tid": 1336753, "ts": 1514656497937.915,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336753, "tid": 1336753,
    "ts": 1514656497941.062, "dur": 3.904,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497943.028, "dur": 1.620,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497949.342, "dur": 125.351,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497950.801, "dur": 2.943,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656497951.288, "dur": 2.318,
    "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": 1336753, "tid": 1336753, "ts": 1514656497951.288,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1336753,
    "ts": 1514656497954.504, "dur": 110.828,
    "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": 1336753, "tid": 1336753, "ts": 1514656497954.504,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656498068.418, "dur": 5.557,
    "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": 1336753, "tid": 1336753, "ts": 1514656498068.418,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336753, "tid": 1336753,
    "ts": 1514656498086.561, "dur": 79.397,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498087.334, "dur": 8.033,
    "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": 1336753, "tid": 1336753, "ts": 1514656498087.334,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336753, "tid": 1336753,
    "ts": 1514656498090.524, "dur": 3.328,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498092.476, "dur": 1.208,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498096.118, "dur": 69.627,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498099.567, "dur": 6.493,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498100.518, "dur": 5.364,
    "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": 1336753, "tid": 1336753, "ts": 1514656498100.518,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1336753,
    "ts": 1514656498106.616, "dur": 52.095,
    "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": 1336753, "tid": 1336753, "ts": 1514656498106.616,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656498160.401, "dur": 4.983,
    "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": 1336753, "tid": 1336753, "ts": 1514656498160.401,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SwiGLULinearFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656498189.948, "dur": 143.966,
    "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": 1336753, "tid": 1336753, "ts": 1514656498189.948,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514656498227.145, "dur": 4.457,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498261.628, "dur": 58.992,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498262.337, "dur": 4.960,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498263.348, "dur": 3.217,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498265.016, "dur": 1.217,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498270.226, "dur": 50.082,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498271.399, "dur": 2.316,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498272.153, "dur": 1.446,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498274.292, "dur": 42.226,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498318.433, "dur": 1.279,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498342.036, "dur": 25.696,
    "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": 1336753, "tid": 1336753, "ts": 1514656498342.036,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656498401.536, "dur": 200.932,
    "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": 1336753, "tid": 1336753, "ts": 1514656498401.536,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1336753,
    "ts": 1514656498420.632, "dur": 3.028,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498421.705, "dur": 1.801,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498430.683, "dur": 6.443,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498433.015, "dur": 3.692,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498442.574, "dur": 3.589,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498582.629, "dur": 3.888,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498583.713, "dur": 2.572,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498622.968, "dur": 90.926,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498623.926, "dur": 7.072,
    "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": 1336753, "tid": 1336753, "ts": 1514656498623.926,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336753, "tid": 1336753,
    "ts": 1514656498625.993, "dur": 3.564,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498627.923, "dur": 1.405,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498631.841, "dur": 81.733,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498635.623, "dur": 4.008,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498636.337, "dur": 3.148,
    "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": 1336753, "tid": 1336753, "ts": 1514656498636.337,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1336753,
    "ts": 1514656498640.359, "dur": 65.363,
    "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": 1336753, "tid": 1336753, "ts": 1514656498640.359,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656498707.805, "dur": 5.077,
    "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": 1336753, "tid": 1336753, "ts": 1514656498707.805,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336753, "tid": 1336753,
    "ts": 1514656498721.783, "dur": 62.044,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498724.531, "dur": 5.835,
    "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": 1336753, "tid": 1336753, "ts": 1514656498724.531,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336753, "tid": 1336753,
    "ts": 1514656498727.203, "dur": 2.188,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498728.546, "dur": 0.695,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498731.045, "dur": 52.562,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498732.050, "dur": 5.030,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498734.752, "dur": 2.192,
    "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": 1336753, "tid": 1336753, "ts": 1514656498734.752,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1336753,
    "ts": 1514656498737.528, "dur": 41.899,
    "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": 1336753, "tid": 1336753, "ts": 1514656498737.528,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656498780.812, "dur": 2.448,
    "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": 1336753, "tid": 1336753, "ts": 1514656498780.812,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336753, "tid": 1336753,
    "ts": 1514656498790.445, "dur": 61.786,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498791.149, "dur": 9.127,
    "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": 1336753, "tid": 1336753, "ts": 1514656498791.149,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336753, "tid": 1336753,
    "ts": 1514656498794.406, "dur": 4.516,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498795.749, "dur": 2.695,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498800.808, "dur": 51.209,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498801.767, "dur": 6.300,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498804.618, "dur": 3.303,
    "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": 1336753, "tid": 1336753, "ts": 1514656498804.618,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1336753,
    "ts": 1514656498808.485, "dur": 37.312,
    "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": 1336753, "tid": 1336753, "ts": 1514656498808.485,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656498847.480, "dur": 4.173,
    "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": 1336753, "tid": 1336753, "ts": 1514656498847.480,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1336753,
    "ts": 1514656498867.332, "dur": 6.445,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498868.549, "dur": 4.824,
    "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": 1336753, "tid": 1336753, "ts": 1514656498868.549,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1336753,
    "ts": 1514656498880.362, "dur": 3.617,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498881.320, "dur": 2.531,
    "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": 1336753, "tid": 1336753, "ts": 1514656498881.320,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1336753,
    "ts": 1514656498887.934, "dur": 3.224,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498889.134, "dur": 1.897,
    "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": 1336753, "tid": 1336753, "ts": 1514656498889.134,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656498917.998, "dur": 192.958,
    "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": 1336753, "tid": 1336753, "ts": 1514656498917.998,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336753, "tid": 1336753,
    "ts": 1514656498938.523, "dur": 10.048,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656498941.581, "dur": 6.611,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499125.327, "dur": 112.256,
    "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": 1336753, "tid": 1336753, "ts": 1514656499125.327,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336753, "tid": 1336753,
    "ts": 1514656499139.267, "dur": 8.847,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499142.694, "dur": 5.048,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499267.391, "dur": 175.032,
    "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": 1336753, "tid": 1336753, "ts": 1514656499267.391,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336753, "tid": 1336753,
    "ts": 1514656499293.604, "dur": 123.033,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499341.859, "dur": 7.317,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499344.845, "dur": 3.814,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499351.857, "dur": 3.949,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499356.879, "dur": 1.301,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499360.637, "dur": 3.043,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499427.822, "dur": 5.364,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499447.931, "dur": 5.370,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499449.219, "dur": 3.944,
    "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": 1336753, "tid": 1336753, "ts": 1514656499449.219,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336753, "tid": 1336753,
    "ts": 1514656499465.022, "dur": 140.791,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499466.608, "dur": 12.759,
    "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": 1336753, "tid": 1336753, "ts": 1514656499466.608,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336753, "tid": 1336753,
    "ts": 1514656499471.631, "dur": 6.513,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499476.153, "dur": 1.689,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499480.640, "dur": 124.757,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499482.285, "dur": 3.379,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499482.958, "dur": 2.558,
    "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": 1336753, "tid": 1336753, "ts": 1514656499482.958,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1336753,
    "ts": 1514656499486.504, "dur": 108.222,
    "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": 1336753, "tid": 1336753, "ts": 1514656499486.504,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656499598.478, "dur": 6.041,
    "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": 1336753, "tid": 1336753, "ts": 1514656499598.478,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656499640.532, "dur": 207.283,
    "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": 1336753, "tid": 1336753, "ts": 1514656499640.532,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1336753,
    "ts": 1514656499658.528, "dur": 3.211,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499659.524, "dur": 2.079,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499665.757, "dur": 5.411,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499669.341, "dur": 1.723,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499670.256, "dur": 0.716,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499678.485, "dur": 7.780,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499680.802, "dur": 4.973,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499692.049, "dur": 3.452,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499699.022, "dur": 3.754,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499825.793, "dur": 3.305,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499826.782, "dur": 2.100,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499831.381, "dur": 2.301,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499832.289, "dur": 1.274,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499865.578, "dur": 111.583,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499875.045, "dur": 9.543,
    "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": 1336753, "tid": 1336753, "ts": 1514656499875.045,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336753, "tid": 1336753,
    "ts": 1514656499879.396, "dur": 4.104,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499881.638, "dur": 1.579,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499885.379, "dur": 91.442,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499886.558, "dur": 7.778,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656499887.365, "dur": 6.805,
    "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": 1336753, "tid": 1336753, "ts": 1514656499887.365,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1336753,
    "ts": 1514656499895.120, "dur": 70.962,
    "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": 1336753, "tid": 1336753, "ts": 1514656499895.120,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656499968.439, "dur": 7.764,
    "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": 1336753, "tid": 1336753, "ts": 1514656499968.439,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336753, "tid": 1336753,
    "ts": 1514656499984.128, "dur": 122.148,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500021.106, "dur": 10.305,
    "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": 1336753, "tid": 1336753, "ts": 1514656500021.106,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336753, "tid": 1336753,
    "ts": 1514656500023.287, "dur": 6.523,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500027.806, "dur": 1.541,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500032.379, "dur": 73.559,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500033.886, "dur": 5.117,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500034.654, "dur": 4.180,
    "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": 1336753, "tid": 1336753, "ts": 1514656500034.654,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1336753,
    "ts": 1514656500042.030, "dur": 59.589,
    "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": 1336753, "tid": 1336753, "ts": 1514656500042.030,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656500103.228, "dur": 2.227,
    "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": 1336753, "tid": 1336753, "ts": 1514656500103.228,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SwiGLULinearFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656500129.026, "dur": 166.990,
    "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": 1336753, "tid": 1336753, "ts": 1514656500129.026,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514656500166.614, "dur": 4.514,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500205.311, "dur": 77.009,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500206.051, "dur": 4.696,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500207.107, "dur": 2.993,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500209.221, "dur": 0.730,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500211.374, "dur": 70.560,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500212.344, "dur": 2.074,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500213.143, "dur": 1.159,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500217.427, "dur": 60.559,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500279.868, "dur": 1.343,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500304.011, "dur": 26.031,
    "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": 1336753, "tid": 1336753, "ts": 1514656500304.011,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656500364.055, "dur": 244.553,
    "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": 1336753, "tid": 1336753, "ts": 1514656500364.055,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1336753,
    "ts": 1514656500381.509, "dur": 2.941,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500382.455, "dur": 1.813,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500396.330, "dur": 9.933,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500401.468, "dur": 4.282,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500433.636, "dur": 6.636,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500591.191, "dur": 4.687,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500592.525, "dur": 3.086,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500628.992, "dur": 93.565,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500630.496, "dur": 10.081,
    "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": 1336753, "tid": 1336753, "ts": 1514656500630.496,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336753, "tid": 1336753,
    "ts": 1514656500632.940, "dur": 6.223,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500637.489, "dur": 1.481,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500641.804, "dur": 80.435,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500643.610, "dur": 3.592,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500644.153, "dur": 2.916,
    "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": 1336753, "tid": 1336753, "ts": 1514656500644.153,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1336753,
    "ts": 1514656500650.421, "dur": 66.578,
    "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": 1336753, "tid": 1336753, "ts": 1514656500650.421,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656500718.938, "dur": 2.592,
    "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": 1336753, "tid": 1336753, "ts": 1514656500718.938,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336753, "tid": 1336753,
    "ts": 1514656500731.008, "dur": 68.400,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500731.765, "dur": 10.776,
    "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": 1336753, "tid": 1336753, "ts": 1514656500731.765,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336753, "tid": 1336753,
    "ts": 1514656500733.570, "dur": 7.562,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500738.109, "dur": 2.627,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500743.269, "dur": 55.909,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500743.957, "dur": 3.295,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500744.851, "dur": 2.280,
    "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": 1336753, "tid": 1336753, "ts": 1514656500744.851,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1336753,
    "ts": 1514656500748.125, "dur": 42.450,
    "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": 1336753, "tid": 1336753, "ts": 1514656500748.125,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656500792.495, "dur": 6.330,
    "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": 1336753, "tid": 1336753, "ts": 1514656500792.495,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336753, "tid": 1336753,
    "ts": 1514656500805.979, "dur": 61.059,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500806.664, "dur": 5.656,
    "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": 1336753, "tid": 1336753, "ts": 1514656500806.664,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336753, "tid": 1336753,
    "ts": 1514656500808.434, "dur": 2.582,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500810.013, "dur": 0.875,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500815.494, "dur": 51.364,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500816.669, "dur": 4.283,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500817.319, "dur": 3.461,
    "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": 1336753, "tid": 1336753, "ts": 1514656500817.319,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1336753,
    "ts": 1514656500821.389, "dur": 38.944,
    "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": 1336753, "tid": 1336753, "ts": 1514656500821.389,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656500862.015, "dur": 4.496,
    "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": 1336753, "tid": 1336753, "ts": 1514656500862.015,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1336753,
    "ts": 1514656500882.084, "dur": 3.305,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500882.639, "dur": 2.632,
    "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": 1336753, "tid": 1336753, "ts": 1514656500882.639,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1336753,
    "ts": 1514656500892.220, "dur": 3.537,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500893.312, "dur": 2.319,
    "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": 1336753, "tid": 1336753, "ts": 1514656500893.312,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1336753,
    "ts": 1514656500902.579, "dur": 5.501,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500903.591, "dur": 4.329,
    "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": 1336753, "tid": 1336753, "ts": 1514656500903.591,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656500935.721, "dur": 183.429,
    "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": 1336753, "tid": 1336753, "ts": 1514656500935.721,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336753, "tid": 1336753,
    "ts": 1514656500954.089, "dur": 8.154,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656500956.830, "dur": 4.868,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501133.218, "dur": 103.258,
    "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": 1336753, "tid": 1336753, "ts": 1514656501133.218,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336753, "tid": 1336753,
    "ts": 1514656501147.616, "dur": 8.645,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501150.041, "dur": 5.843,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501262.766, "dur": 176.642,
    "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": 1336753, "tid": 1336753, "ts": 1514656501262.766,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336753, "tid": 1336753,
    "ts": 1514656501289.646, "dur": 125.084,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501340.258, "dur": 7.090,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501342.382, "dur": 4.495,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501349.881, "dur": 4.460,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501355.681, "dur": 3.675,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501361.703, "dur": 3.130,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501425.881, "dur": 4.673,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501444.518, "dur": 5.241,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501445.665, "dur": 3.950,
    "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": 1336753, "tid": 1336753, "ts": 1514656501445.665,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336753, "tid": 1336753,
    "ts": 1514656501461.761, "dur": 127.715,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501465.119, "dur": 8.283,
    "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": 1336753, "tid": 1336753, "ts": 1514656501465.119,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336753, "tid": 1336753,
    "ts": 1514656501467.894, "dur": 4.374,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501470.736, "dur": 1.321,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501474.753, "dur": 114.320,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501476.395, "dur": 5.892,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501479.544, "dur": 2.551,
    "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": 1336753, "tid": 1336753, "ts": 1514656501479.544,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1336753,
    "ts": 1514656501483.160, "dur": 96.643,
    "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": 1336753, "tid": 1336753, "ts": 1514656501483.160,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656501583.767, "dur": 4.422,
    "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": 1336753, "tid": 1336753, "ts": 1514656501583.767,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656501624.408, "dur": 219.040,
    "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": 1336753, "tid": 1336753, "ts": 1514656501624.408,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336753, "tid": 1336753,
    "ts": 1514656501644.620, "dur": 3.725,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501645.943, "dur": 2.060,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501651.856, "dur": 3.868,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501653.408, "dur": 2.210,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501654.434, "dur": 1.019,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501663.096, "dur": 9.228,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501666.903, "dur": 5.049,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501678.432, "dur": 3.460,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501685.523, "dur": 5.216,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501820.040, "dur": 3.383,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501820.998, "dur": 2.054,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501826.070, "dur": 2.640,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501826.970, "dur": 1.631,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501862.045, "dur": 96.573,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501863.505, "dur": 10.556,
    "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": 1336753, "tid": 1336753, "ts": 1514656501863.505,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336753, "tid": 1336753,
    "ts": 1514656501866.376, "dur": 6.598,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501868.906, "dur": 3.743,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501875.098, "dur": 83.147,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501877.121, "dur": 6.053,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501879.977, "dur": 3.053,
    "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": 1336753, "tid": 1336753, "ts": 1514656501879.977,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1336753,
    "ts": 1514656501884.165, "dur": 66.294,
    "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": 1336753, "tid": 1336753, "ts": 1514656501884.165,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656501952.512, "dur": 5.071,
    "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": 1336753, "tid": 1336753, "ts": 1514656501952.512,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336753, "tid": 1336753,
    "ts": 1514656501966.359, "dur": 121.749,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501967.243, "dur": 10.049,
    "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": 1336753, "tid": 1336753, "ts": 1514656501967.243,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336753, "tid": 1336753,
    "ts": 1514656501972.937, "dur": 3.177,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501974.810, "dur": 1.115,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501978.126, "dur": 109.782,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501978.813, "dur": 43.218,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656501981.962, "dur": 39.609,
    "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": 1336753, "tid": 1336753, "ts": 1514656501981.962,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336753, "tid": 1336753,
    "ts": 1514656502023.329, "dur": 59.028,
    "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": 1336753, "tid": 1336753, "ts": 1514656502023.329,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656502084.685, "dur": 2.741,
    "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": 1336753, "tid": 1336753, "ts": 1514656502084.685,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SwiGLULinearFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656502111.398, "dur": 150.725,
    "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": 1336753, "tid": 1336753, "ts": 1514656502111.398,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336753, "tid": 1336753,
    "ts": 1514656502151.148, "dur": 4.802,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502187.181, "dur": 62.585,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502188.126, "dur": 6.928,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502189.681, "dur": 4.465,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502191.300, "dur": 2.617,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502195.816, "dur": 53.484,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502197.307, "dur": 4.892,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502200.514, "dur": 1.565,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502202.690, "dur": 42.541,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502247.505, "dur": 1.141,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502270.781, "dur": 25.205,
    "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": 1336753, "tid": 1336753, "ts": 1514656502270.781,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::stack", "pid": 1336753, "tid": 1336753,
    "ts": 1514656502314.684, "dur": 45.208,
    "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": 1336753, "tid": 1336753, "ts": 1514656502314.684,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 1336753, "tid": 1336753,
    "ts": 1514656502324.441, "dur": 30.556,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502356.558, "dur": 1.937,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502398.817, "dur": 51.991,
    "args": {
      "External id": 2480937,"Record function id": 0, "Ev Idx": 8640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 2/0", "pid": 1336753, "tid": 1336753,
    "ts": 1514656502451.968, "dur": 225.835,
    "args": {
      "External id": 2480938,"Record function id": 0, "Ev Idx": 8641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336753, "tid": 1336753,
    "ts": 1514656502490.341, "dur": 178.775,
    "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": 1336753, "tid": 1336753, "ts": 1514656502490.341,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336753, "tid": 1336753,
    "ts": 1514656502584.817, "dur": 44.098,
    "args": {
      "External id": 2480940,"kernel_hash": "cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/ml/cmljljlwjilefo6ueuqn5gaz5uvsdhhmauseg2vxhp3egffs57dc.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": 1336753, "tid": 1336753,
    "ts": 1514656502767.096, "dur": 41.587,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502769.888, "dur": 5.512,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502778.544, "dur": 29.806,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502781.762, "dur": 26.088,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502813.138, "dur": 20.216,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502814.282, "dur": 3.180,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502819.969, "dur": 13.130,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502820.680, "dur": 11.956,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502835.846, "dur": 17.152,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502836.708, "dur": 3.043,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502840.315, "dur": 12.425,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502841.223, "dur": 11.144,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502863.307, "dur": 0.854,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502872.337, "dur": 11.240,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502879.904, "dur": 2.272,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502889.734, "dur": 7.427,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502893.770, "dur": 1.662,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502898.542, "dur": 3.641,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502900.661, "dur": 0.849,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502903.499, "dur": 3.422,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502905.451, "dur": 0.967,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502909.798, "dur": 3.620,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502912.101, "dur": 0.764,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502914.429, "dur": 4.088,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502916.478, "dur": 1.253,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502919.694, "dur": 3.323,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502921.556, "dur": 0.971,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502926.464, "dur": 5.835,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502930.744, "dur": 0.753,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502936.077, "dur": 2.975,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502937.970, "dur": 0.533,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502945.053, "dur": 7.263,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502950.291, "dur": 0.910,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502953.484, "dur": 2.735,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502955.058, "dur": 0.619,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502958.613, "dur": 8.474,
    "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": 1336753, "tid": 1336753, "ts": 1514656502958.613,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1336753,
    "ts": 1514656502964.565, "dur": 0.941,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502968.178, "dur": 5.763,
    "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": 1336753, "tid": 1336753, "ts": 1514656502968.178,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1336753,
    "ts": 1514656502972.382, "dur": 0.847,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656502975.011, "dur": 8.131,
    "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": 1336753, "tid": 1336753, "ts": 1514656502975.011,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1336753,
    "ts": 1514656502981.373, "dur": 0.882,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503018.712, "dur": 9.352,
    "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": 1336753, "tid": 1336753, "ts": 1514656503018.712,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1336753,
    "ts": 1514656503023.177, "dur": 3.233,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503034.111, "dur": 41.882,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503036.483, "dur": 39.313,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503038.922, "dur": 7.029,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503041.181, "dur": 4.158,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503047.784, "dur": 27.503,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503102.119, "dur": 4.845,
    "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": 1336753, "tid": 1336753, "ts": 1514656503102.119,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656503109.090, "dur": 1.485,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503138.756, "dur": 72435.498,
    "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": 1336753, "tid": 1336753, "ts": 1514656503138.756,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 1336753, "tid": 1336753,
    "ts": 1514656503153.241, "dur": 32.370,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503154.451, "dur": 30.946,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503155.886, "dur": 7.369,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503157.227, "dur": 5.572,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503164.179, "dur": 20.425,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503203.897, "dur": 26.632,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503205.295, "dur": 6.390,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503207.834, "dur": 3.537,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503212.501, "dur": 17.806,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503213.663, "dur": 16.293,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503233.777, "dur": 42.208,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503235.066, "dur": 6.946,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503238.688, "dur": 3.085,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503261.955, "dur": 13.814,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503262.769, "dur": 12.488,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503281.491, "dur": 20.132,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503282.822, "dur": 4.928,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503288.587, "dur": 12.739,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503289.047, "dur": 11.974,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503311.315, "dur": 28.838,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503344.073, "dur": 61.114,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503347.764, "dur": 57.008,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503354.109, "dur": 1.054,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503357.175, "dur": 27.126,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503359.035, "dur": 25.027,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503363.927, "dur": 2.978,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503367.642, "dur": 15.989,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503411.493, "dur": 65998.356,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656503413.966, "dur": 65994.936,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569421.771, "dur": 7.898,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569426.818, "dur": 1.009,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569434.920, "dur": 125.315,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569439.030, "dur": 10.098,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569441.453, "dur": 6.781,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569446.990, "dur": 0.963,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569450.461, "dur": 109.194,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569452.320, "dur": 105.854,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569564.998, "dur": 5.458,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569567.927, "dur": 0.765,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569581.101, "dur": 4.238,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569597.328, "dur": 7.668,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569599.836, "dur": 4.883,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569751.431, "dur": 282.183,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569756.759, "dur": 2.654,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569762.024, "dur": 271.046,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569767.624, "dur": 0.405,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569771.247, "dur": 29.343,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569803.248, "dur": 5.814,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569807.818, "dur": 0.773,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569811.324, "dur": 28.449,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569812.843, "dur": 1.486,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569816.283, "dur": 23.147,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569820.724, "dur": 3.213,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569842.463, "dur": 25.449,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569870.259, "dur": 22.924,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569899.102, "dur": 18.381,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569919.658, "dur": 16.076,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569938.336, "dur": 24.484,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569940.183, "dur": 1.907,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569944.511, "dur": 0.897,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656569965.399, "dur": 15.586,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570015.700, "dur": 15.540,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570042.732, "dur": 2.664,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570053.035, "dur": 4.851,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570055.878, "dur": 1.078,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570147.960, "dur": 77.934,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570231.337, "dur": 5.175,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570234.014, "dur": 1.105,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570238.408, "dur": 33.720,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570280.265, "dur": 7.115,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570282.182, "dur": 4.449,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570284.493, "dur": 1.888,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570291.734, "dur": 50.193,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570293.519, "dur": 47.771,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570347.056, "dur": 19.809,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570372.834, "dur": 6.105,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570377.015, "dur": 0.873,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570383.212, "dur": 53.485,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570384.448, "dur": 4.150,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570385.331, "dur": 2.657,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570386.921, "dur": 0.912,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570389.323, "dur": 46.821,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570392.269, "dur": 43.238,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570440.745, "dur": 4.274,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570442.815, "dur": 1.177,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570451.700, "dur": 1.721,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570461.595, "dur": 12.156,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570466.386, "dur": 7.064,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570591.218, "dur": 207.742,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570594.318, "dur": 3.451,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570599.364, "dur": 199.055,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570601.061, "dur": 0.293,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570602.971, "dur": 27.747,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570632.692, "dur": 4.467,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570635.391, "dur": 1.298,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570638.244, "dur": 28.755,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570641.476, "dur": 1.720,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570644.552, "dur": 22.049,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570650.385, "dur": 3.044,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570668.942, "dur": 24.347,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570694.765, "dur": 14.326,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570712.763, "dur": 15.342,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570729.401, "dur": 13.895,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570745.595, "dur": 24.176,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570747.364, "dur": 1.436,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570753.263, "dur": 0.712,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570771.334, "dur": 12.692,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570785.282, "dur": 12.048,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570806.149, "dur": 2.259,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570817.671, "dur": 3.633,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570820.133, "dur": 0.336,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570891.553, "dur": 57.586,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570954.208, "dur": 7.317,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570959.151, "dur": 1.242,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656570962.796, "dur": 62.796,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571032.530, "dur": 6.453,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571034.270, "dur": 3.874,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571036.736, "dur": 1.171,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571041.969, "dur": 49.691,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571045.298, "dur": 45.556,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571095.950, "dur": 17.154,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571119.360, "dur": 3.840,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571121.277, "dur": 1.030,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571127.502, "dur": 56.952,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571128.802, "dur": 7.836,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571129.713, "dur": 6.293,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571133.178, "dur": 2.595,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571137.497, "dur": 46.352,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571138.396, "dur": 44.699,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571188.486, "dur": 4.063,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571190.277, "dur": 1.140,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571199.013, "dur": 1.368,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571208.347, "dur": 8.926,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571212.659, "dur": 4.389,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571310.041, "dur": 322.491,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571313.321, "dur": 2.270,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571316.902, "dur": 315.038,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571318.625, "dur": 0.330,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571322.855, "dur": 22.672,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571347.321, "dur": 3.376,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571349.509, "dur": 0.876,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571351.867, "dur": 27.548,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571355.134, "dur": 3.498,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571359.773, "dur": 19.239,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571362.467, "dur": 2.744,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571380.747, "dur": 24.424,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571406.585, "dur": 25.472,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571435.529, "dur": 40.472,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571477.311, "dur": 42.409,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571521.692, "dur": 45.471,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571523.521, "dur": 1.712,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571547.099, "dur": 0.957,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571569.069, "dur": 33.859,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571604.206, "dur": 26.693,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571640.702, "dur": 2.110,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571652.586, "dur": 3.626,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571654.949, "dur": 0.346,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571727.151, "dur": 53.691,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571787.649, "dur": 5.103,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571790.250, "dur": 1.435,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571794.212, "dur": 24.884,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571912.726, "dur": 9.141,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571914.857, "dur": 5.932,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571917.516, "dur": 3.059,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571927.332, "dur": 46.621,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571928.377, "dur": 44.796,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656571978.308, "dur": 65.639,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572052.793, "dur": 5.010,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572055.537, "dur": 0.936,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572062.647, "dur": 56.257,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572063.493, "dur": 6.746,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572066.860, "dur": 2.821,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572068.751, "dur": 0.569,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572070.872, "dur": 47.591,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572071.553, "dur": 46.159,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572122.864, "dur": 3.920,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572124.672, "dur": 0.815,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572136.391, "dur": 1.639,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572146.835, "dur": 7.276,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572148.827, "dur": 5.033,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572251.455, "dur": 197.200,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572254.041, "dur": 2.171,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572259.758, "dur": 188.367,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572263.344, "dur": 0.384,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572264.910, "dur": 24.008,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572290.485, "dur": 5.161,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572292.650, "dur": 2.649,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572296.635, "dur": 22.928,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572297.599, "dur": 1.489,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572300.315, "dur": 18.924,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572303.427, "dur": 2.551,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572321.337, "dur": 24.191,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572346.965, "dur": 14.613,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572366.584, "dur": 16.127,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572384.363, "dur": 13.434,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572399.536, "dur": 20.650,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572401.734, "dur": 1.297,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572404.951, "dur": 0.670,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572421.467, "dur": 12.944,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572435.483, "dur": 11.759,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572457.260, "dur": 1.374,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572466.202, "dur": 3.198,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572468.168, "dur": 0.378,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572549.733, "dur": 58.133,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572613.564, "dur": 7.989,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572616.921, "dur": 3.035,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572623.245, "dur": 28.180,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572656.430, "dur": 7.799,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572660.392, "dur": 3.028,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572662.355, "dur": 0.925,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572666.948, "dur": 40.588,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572667.747, "dur": 39.271,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572711.811, "dur": 16.584,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572733.646, "dur": 6.133,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572738.306, "dur": 0.607,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572743.895, "dur": 47.335,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572744.879, "dur": 3.424,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572745.831, "dur": 1.958,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572747.126, "dur": 0.547,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572749.152, "dur": 41.655,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572749.712, "dur": 40.541,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572794.847, "dur": 4.051,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572796.783, "dur": 1.028,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572807.190, "dur": 1.503,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572816.895, "dur": 6.798,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572818.811, "dur": 4.647,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572904.135, "dur": 239.131,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572906.277, "dur": 3.890,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572914.235, "dur": 228.490,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572915.621, "dur": 0.285,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572916.848, "dur": 22.354,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572940.903, "dur": 5.284,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572945.236, "dur": 0.627,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572947.109, "dur": 21.618,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572948.711, "dur": 1.379,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572951.209, "dur": 17.234,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572953.401, "dur": 2.681,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656572972.321, "dur": 58.734,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573033.777, "dur": 16.339,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573052.901, "dur": 14.319,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573068.557, "dur": 14.301,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573084.672, "dur": 23.863,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573089.178, "dur": 1.378,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573092.518, "dur": 1.180,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573109.910, "dur": 16.099,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573129.426, "dur": 12.194,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573151.120, "dur": 2.140,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573162.152, "dur": 3.848,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573164.409, "dur": 0.484,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573234.812, "dur": 53.811,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573293.588, "dur": 4.886,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573296.104, "dur": 1.357,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573299.828, "dur": 23.785,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573329.946, "dur": 4.959,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573331.260, "dur": 3.103,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573333.074, "dur": 1.073,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573337.419, "dur": 38.546,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573338.581, "dur": 36.881,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573379.140, "dur": 14.861,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573399.041, "dur": 6.242,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573403.384, "dur": 1.069,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573409.259, "dur": 49.273,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573410.596, "dur": 3.846,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573411.428, "dur": 2.508,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573412.790, "dur": 1.031,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573415.213, "dur": 42.966,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573418.146, "dur": 39.573,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573462.426, "dur": 4.046,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573464.314, "dur": 1.063,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573480.579, "dur": 1.625,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573490.600, "dur": 10.775,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573494.374, "dur": 6.714,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573599.442, "dur": 234.924,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573601.590, "dur": 3.087,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573606.727, "dur": 227.186,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573641.307, "dur": 0.530,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573646.079, "dur": 25.062,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573672.805, "dur": 6.040,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573675.573, "dur": 2.902,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573679.834, "dur": 25.915,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573683.857, "dur": 1.916,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573686.900, "dur": 18.440,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573689.469, "dur": 2.660,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573707.037, "dur": 22.360,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573730.990, "dur": 14.313,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573747.818, "dur": 16.774,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573765.956, "dur": 12.477,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573780.109, "dur": 25.193,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573781.983, "dur": 1.865,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573788.069, "dur": 2.718,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573806.655, "dur": 12.976,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573820.683, "dur": 12.403,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573841.324, "dur": 2.010,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573852.995, "dur": 3.255,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573855.005, "dur": 0.401,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573918.268, "dur": 57.556,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656573982.952, "dur": 40.084,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574019.585, "dur": 1.345,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574024.551, "dur": 28.870,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574058.650, "dur": 6.129,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574060.125, "dur": 3.789,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574062.523, "dur": 1.224,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574070.250, "dur": 44.985,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574071.145, "dur": 43.551,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574119.437, "dur": 15.699,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574140.891, "dur": 4.096,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574143.160, "dur": 0.895,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574148.945, "dur": 49.086,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574150.122, "dur": 5.400,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574153.157, "dur": 1.811,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574154.227, "dur": 0.623,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574156.063, "dur": 41.636,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574156.710, "dur": 40.309,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574202.093, "dur": 4.179,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574204.107, "dur": 1.062,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574212.094, "dur": 1.273,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574222.998, "dur": 6.264,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574224.687, "dur": 4.337,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574311.402, "dur": 189.584,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574315.597, "dur": 2.267,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574319.579, "dur": 180.907,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574320.862, "dur": 0.311,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574322.322, "dur": 22.374,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574348.985, "dur": 5.572,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574353.277, "dur": 0.986,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574355.362, "dur": 22.094,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574356.385, "dur": 1.702,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574359.164, "dur": 17.962,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574361.534, "dur": 2.545,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574378.812, "dur": 20.639,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574401.008, "dur": 17.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574420.750, "dur": 14.484,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574436.522, "dur": 13.179,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574453.263, "dur": 20.685,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574455.089, "dur": 1.464,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574458.756, "dur": 0.910,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574475.148, "dur": 12.824,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574488.886, "dur": 10.643,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574506.539, "dur": 1.374,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574515.236, "dur": 2.802,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574516.985, "dur": 0.307,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574601.709, "dur": 54.331,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574660.964, "dur": 5.827,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574664.461, "dur": 0.908,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574668.237, "dur": 27.169,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574700.455, "dur": 7.474,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574701.842, "dur": 5.529,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574706.143, "dur": 1.045,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574710.502, "dur": 40.755,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574711.506, "dur": 39.117,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574754.621, "dur": 16.125,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574775.803, "dur": 3.973,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574777.906, "dur": 1.065,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574783.366, "dur": 49.297,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574786.459, "dur": 3.502,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574787.386, "dur": 2.039,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574788.676, "dur": 0.623,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574790.469, "dur": 41.745,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574791.106, "dur": 40.492,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574836.316, "dur": 4.314,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574838.522, "dur": 1.069,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574848.456, "dur": 1.425,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574857.209, "dur": 8.153,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574859.035, "dur": 6.060,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574948.232, "dur": 237.662,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574950.530, "dur": 2.346,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574956.823, "dur": 228.617,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574958.239, "dur": 0.282,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656574959.330, "dur": 58.156,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575020.890, "dur": 3.804,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575023.203, "dur": 1.117,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575025.693, "dur": 28.406,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575026.840, "dur": 2.386,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575030.341, "dur": 23.433,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575035.675, "dur": 2.933,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575057.858, "dur": 22.567,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575082.001, "dur": 14.640,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575099.692, "dur": 16.869,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575117.932, "dur": 13.248,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575132.791, "dur": 21.297,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575134.642, "dur": 1.189,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575137.604, "dur": 0.996,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575155.583, "dur": 12.581,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575172.018, "dur": 12.422,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575193.867, "dur": 2.149,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575205.425, "dur": 3.440,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575207.633, "dur": 0.359,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575274.226, "dur": 54.752,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575333.434, "dur": 5.025,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575336.202, "dur": 1.209,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575339.626, "dur": 24.936,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575371.188, "dur": 5.282,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575372.686, "dur": 2.928,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575374.329, "dur": 1.048,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575378.801, "dur": 39.158,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575379.988, "dur": 37.342,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575421.501, "dur": 15.445,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575441.021, "dur": 31.295,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575446.133, "dur": 25.793,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575453.676, "dur": 1.078,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575477.674, "dur": 27.278,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575479.518, "dur": 25.223,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575484.074, "dur": 4.223,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575489.451, "dur": 14.847,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575520.466, "dur": 5.300,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575522.413, "dur": 3.086,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575542.714, "dur": 5.254,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575546.229, "dur": 1.460,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575591.319, "dur": 26.554,
    "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": 1336753, "tid": 1336753, "ts": 1514656575591.319,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336753, "tid": 1336753,
    "ts": 1514656575624.466, "dur": 13.603,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575635.237, "dur": 1.203,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575640.572, "dur": 6.884,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575645.269, "dur": 0.874,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575651.236, "dur": 3.127,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575652.663, "dur": 1.118,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575657.943, "dur": 6.147,
    "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": 1336753, "tid": 1336753, "ts": 1514656575657.943,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1336753,
    "ts": 1514656575661.689, "dur": 1.076,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575665.066, "dur": 7.847,
    "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": 1336753, "tid": 1336753, "ts": 1514656575665.066,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1336753,
    "ts": 1514656575671.269, "dur": 0.864,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575673.959, "dur": 5.078,
    "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": 1336753, "tid": 1336753, "ts": 1514656575673.959,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1336753,
    "ts": 1514656575677.506, "dur": 0.652,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575680.026, "dur": 4.776,
    "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": 1336753, "tid": 1336753, "ts": 1514656575680.026,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1336753,
    "ts": 1514656575682.968, "dur": 1.065,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575688.630, "dur": 33.372,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575690.196, "dur": 31.564,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575692.983, "dur": 9.158,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575697.782, "dur": 3.849,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575702.973, "dur": 18.298,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575747.209, "dur": 3.936,
    "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": 1336753, "tid": 1336753, "ts": 1514656575747.209,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656575753.287, "dur": 1.489,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575785.741, "dur": 42949.087,
    "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": 1336753, "tid": 1336753, "ts": 1514656575785.741,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 1336753, "tid": 1336753,
    "ts": 1514656575799.452, "dur": 31.031,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575800.149, "dur": 30.090,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575801.743, "dur": 9.981,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575807.884, "dur": 3.437,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575812.408, "dur": 17.281,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575845.214, "dur": 26.405,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575846.655, "dur": 6.281,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575848.529, "dur": 4.055,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575854.212, "dur": 17.196,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575855.882, "dur": 15.074,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575874.926, "dur": 23.051,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575875.668, "dur": 4.663,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575876.824, "dur": 3.261,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575880.847, "dur": 16.953,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575883.966, "dur": 13.451,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575903.371, "dur": 19.938,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575905.100, "dur": 3.690,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575909.724, "dur": 13.307,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575910.329, "dur": 12.387,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575927.928, "dur": 20.835,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656575951.656, "dur": 144.283,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656576032.551, "dur": 62.742,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656576042.524, "dur": 1.200,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656576045.546, "dur": 32.797,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656576049.884, "dur": 28.208,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656576052.367, "dur": 3.855,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656576057.086, "dur": 20.552,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656576100.805, "dur": 36770.384,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656576102.579, "dur": 36767.624,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656612883.037, "dur": 7.194,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656612887.058, "dur": 1.278,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656612895.706, "dur": 131.372,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656612899.696, "dur": 6.754,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656612902.122, "dur": 3.568,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656612904.224, "dur": 1.189,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656612907.699, "dur": 118.551,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656612909.708, "dur": 115.170,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613031.950, "dur": 5.136,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613034.830, "dur": 0.788,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613047.430, "dur": 2.571,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613059.679, "dur": 6.890,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613061.704, "dur": 4.563,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613193.422, "dur": 217.170,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613197.003, "dur": 2.358,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613200.778, "dur": 209.364,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613204.841, "dur": 0.454,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613207.022, "dur": 30.480,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613239.316, "dur": 4.219,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613241.903, "dur": 1.283,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613244.994, "dur": 27.693,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613246.026, "dur": 1.605,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613248.914, "dur": 23.454,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613254.809, "dur": 3.386,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613274.309, "dur": 23.546,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613299.716, "dur": 15.769,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613321.164, "dur": 17.157,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613339.885, "dur": 14.167,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613356.435, "dur": 22.390,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613358.284, "dur": 1.883,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613362.389, "dur": 0.877,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613380.463, "dur": 12.957,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613397.025, "dur": 11.812,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613417.526, "dur": 1.717,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613425.983, "dur": 4.095,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613428.428, "dur": 0.689,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613504.160, "dur": 82.375,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613593.605, "dur": 8.047,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613596.793, "dur": 2.120,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613603.230, "dur": 32.450,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613644.137, "dur": 5.913,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613645.884, "dur": 3.539,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613648.141, "dur": 1.045,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613653.278, "dur": 46.849,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613654.484, "dur": 45.011,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613704.987, "dur": 18.082,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613729.174, "dur": 6.134,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613733.571, "dur": 0.844,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613739.707, "dur": 59.043,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613740.755, "dur": 5.392,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613741.421, "dur": 4.172,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613742.828, "dur": 2.583,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613746.754, "dur": 51.589,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613754.994, "dur": 42.571,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613802.407, "dur": 3.766,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613804.441, "dur": 0.672,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613812.856, "dur": 1.765,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613823.005, "dur": 8.715,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613827.144, "dur": 4.336,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613922.192, "dur": 256.253,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613926.592, "dur": 2.259,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613930.326, "dur": 247.569,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613931.777, "dur": 0.291,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613933.548, "dur": 23.766,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613974.411, "dur": 38.645,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656613980.475, "dur": 31.638,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614017.401, "dur": 26.818,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614018.716, "dur": 2.649,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614022.645, "dur": 21.262,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614026.072, "dur": 3.015,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614045.827, "dur": 24.530,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614071.919, "dur": 16.870,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614091.751, "dur": 15.324,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614108.416, "dur": 13.983,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614124.226, "dur": 24.792,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614128.146, "dur": 1.233,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614131.880, "dur": 0.967,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614150.267, "dur": 13.237,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614164.817, "dur": 12.090,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614187.575, "dur": 2.457,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614199.985, "dur": 3.997,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614202.704, "dur": 0.361,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614277.835, "dur": 58.506,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614343.600, "dur": 4.858,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614346.504, "dur": 0.704,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614349.714, "dur": 25.974,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614380.605, "dur": 5.372,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614382.048, "dur": 3.262,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614383.908, "dur": 1.255,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614390.545, "dur": 40.287,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614391.903, "dur": 38.229,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614435.005, "dur": 15.302,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614455.669, "dur": 3.492,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614457.757, "dur": 0.576,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614463.020, "dur": 49.269,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614464.069, "dur": 6.024,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614467.111, "dur": 2.412,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614468.436, "dur": 0.910,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614470.875, "dur": 40.895,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614471.366, "dur": 39.869,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614516.029, "dur": 3.778,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614517.749, "dur": 1.020,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614524.830, "dur": 1.387,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614553.260, "dur": 10.165,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614555.471, "dur": 7.484,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614654.503, "dur": 192.780,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614657.024, "dur": 2.279,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614660.614, "dur": 186.040,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614662.106, "dur": 0.240,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614663.852, "dur": 25.288,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614690.535, "dur": 3.584,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614692.910, "dur": 0.884,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614697.835, "dur": 25.144,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614698.530, "dur": 1.584,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614701.023, "dur": 21.672,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614706.287, "dur": 2.858,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614724.268, "dur": 21.264,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614746.760, "dur": 14.569,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614764.558, "dur": 14.603,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614780.550, "dur": 12.743,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614795.127, "dur": 24.180,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614798.868, "dur": 1.947,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614803.081, "dur": 1.178,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614820.501, "dur": 12.691,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614834.338, "dur": 11.375,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614853.300, "dur": 2.012,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614864.080, "dur": 3.575,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614866.551, "dur": 0.344,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656614933.997, "dur": 90.810,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615032.081, "dur": 6.470,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615035.607, "dur": 1.466,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615039.931, "dur": 28.586,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615073.499, "dur": 7.555,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615075.214, "dur": 5.231,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615079.363, "dur": 0.901,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615083.997, "dur": 62.002,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615085.161, "dur": 60.134,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615150.077, "dur": 23.612,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615179.087, "dur": 3.770,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615181.124, "dur": 0.942,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615186.687, "dur": 53.037,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615189.789, "dur": 6.602,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615190.388, "dur": 5.367,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615191.926, "dur": 3.637,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615197.192, "dur": 42.031,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615197.711, "dur": 40.869,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615243.087, "dur": 3.880,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615245.361, "dur": 0.506,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615255.656, "dur": 1.488,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615264.528, "dur": 6.177,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615266.589, "dur": 3.870,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615351.643, "dur": 213.592,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615354.184, "dur": 2.428,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615360.231, "dur": 204.488,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615361.234, "dur": 0.269,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615365.022, "dur": 23.123,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615389.698, "dur": 3.697,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615392.094, "dur": 1.008,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615394.542, "dur": 24.551,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615395.462, "dur": 3.661,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615400.611, "dur": 18.214,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615403.417, "dur": 2.853,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615422.792, "dur": 22.349,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615446.231, "dur": 13.435,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615462.605, "dur": 15.005,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615479.006, "dur": 13.025,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615493.625, "dur": 21.023,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615495.233, "dur": 1.426,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615498.711, "dur": 1.056,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615515.938, "dur": 29.420,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615549.491, "dur": 13.827,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615572.492, "dur": 1.911,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615582.335, "dur": 4.414,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615585.098, "dur": 0.766,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615651.124, "dur": 52.795,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615708.660, "dur": 4.332,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615710.968, "dur": 0.946,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615714.212, "dur": 24.470,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615745.280, "dur": 7.745,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615746.594, "dur": 5.822,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615748.708, "dur": 3.500,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615755.536, "dur": 39.456,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615756.542, "dur": 37.792,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615798.848, "dur": 14.872,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615818.363, "dur": 5.309,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615822.096, "dur": 0.759,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615827.167, "dur": 49.303,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615827.966, "dur": 4.040,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615828.882, "dur": 2.545,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615830.360, "dur": 0.938,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615832.713, "dur": 43.344,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615835.459, "dur": 40.004,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615880.139, "dur": 3.733,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615881.978, "dur": 0.931,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615888.982, "dur": 1.138,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615897.151, "dur": 6.802,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656615898.982, "dur": 4.691,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616017.980, "dur": 286.009,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616020.959, "dur": 3.713,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616028.254, "dur": 275.154,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616029.750, "dur": 0.273,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616031.132, "dur": 24.606,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616057.194, "dur": 6.273,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616059.719, "dur": 3.368,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616064.521, "dur": 24.412,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616065.626, "dur": 1.684,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616068.421, "dur": 20.236,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616073.446, "dur": 2.848,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616090.149, "dur": 22.736,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616114.225, "dur": 15.093,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616131.784, "dur": 14.105,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616147.164, "dur": 13.406,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616162.049, "dur": 111.128,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616252.336, "dur": 2.034,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616256.578, "dur": 0.933,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616277.398, "dur": 13.308,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616291.744, "dur": 10.954,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616311.989, "dur": 2.055,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616323.371, "dur": 3.150,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616325.461, "dur": 0.361,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616399.226, "dur": 60.206,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616464.113, "dur": 7.525,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616469.520, "dur": 0.996,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616472.915, "dur": 24.990,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616502.289, "dur": 5.394,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616503.859, "dur": 3.177,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616505.541, "dur": 1.278,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616510.486, "dur": 60.197,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616511.716, "dur": 57.981,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616579.054, "dur": 17.367,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616603.333, "dur": 4.452,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616605.800, "dur": 0.961,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616611.888, "dur": 55.763,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616612.847, "dur": 6.221,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616613.668, "dur": 4.863,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616617.562, "dur": 0.830,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616619.668, "dur": 47.651,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616620.364, "dur": 46.426,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616671.091, "dur": 4.400,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616673.149, "dur": 1.125,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616681.806, "dur": 1.624,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616693.398, "dur": 6.471,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616695.055, "dur": 4.516,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616787.285, "dur": 245.434,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616791.672, "dur": 2.438,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616795.487, "dur": 236.749,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616796.586, "dur": 0.632,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616798.157, "dur": 22.150,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616821.927, "dur": 5.363,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616825.879, "dur": 1.135,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616830.481, "dur": 22.740,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616831.616, "dur": 1.665,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616834.376, "dur": 18.545,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616837.170, "dur": 2.825,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616854.561, "dur": 22.307,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616878.203, "dur": 16.117,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616896.986, "dur": 15.572,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616914.162, "dur": 13.620,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616929.979, "dur": 24.351,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616933.838, "dur": 1.680,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616937.880, "dur": 0.980,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616955.331, "dur": 12.733,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656616969.291, "dur": 60.652,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617041.712, "dur": 2.759,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617054.695, "dur": 4.714,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617057.832, "dur": 0.682,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617130.943, "dur": 56.223,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617191.961, "dur": 4.639,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617194.465, "dur": 1.174,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617197.995, "dur": 26.373,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617228.729, "dur": 8.648,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617230.347, "dur": 6.453,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617234.980, "dur": 1.600,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617239.968, "dur": 39.394,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617241.178, "dur": 37.465,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617283.389, "dur": 15.647,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617304.637, "dur": 3.412,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617306.478, "dur": 0.722,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617311.889, "dur": 49.635,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617315.332, "dur": 4.161,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617316.283, "dur": 2.699,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617318.015, "dur": 0.797,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617320.013, "dur": 41.111,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617320.885, "dur": 39.592,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617364.956, "dur": 3.980,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617367.149, "dur": 0.699,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617374.153, "dur": 1.402,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617385.139, "dur": 8.513,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617387.129, "dur": 6.232,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617470.790, "dur": 228.085,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617473.622, "dur": 2.636,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617477.839, "dur": 220.393,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617479.231, "dur": 0.298,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617483.088, "dur": 30.126,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617514.892, "dur": 3.619,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617517.159, "dur": 1.071,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617541.742, "dur": 29.311,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617543.522, "dur": 4.452,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617549.522, "dur": 20.953,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617553.036, "dur": 2.858,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617572.191, "dur": 21.848,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617595.498, "dur": 15.112,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617616.006, "dur": 14.840,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617632.167, "dur": 13.710,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617647.708, "dur": 22.321,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617649.971, "dur": 1.138,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617653.610, "dur": 1.274,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617671.475, "dur": 13.380,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617685.964, "dur": 11.456,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617708.735, "dur": 1.810,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617718.677, "dur": 4.005,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617721.313, "dur": 0.527,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617786.650, "dur": 50.395,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617841.513, "dur": 4.602,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617843.947, "dur": 1.087,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617850.316, "dur": 22.391,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617879.698, "dur": 10.184,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617881.025, "dur": 8.225,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617885.745, "dur": 3.253,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617892.428, "dur": 40.319,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617893.583, "dur": 38.540,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617936.224, "dur": 14.893,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617956.495, "dur": 5.417,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617960.505, "dur": 0.620,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617965.954, "dur": 87.641,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617966.934, "dur": 4.264,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617967.937, "dur": 2.733,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617969.654, "dur": 0.908,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617971.894, "dur": 81.168,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656617975.045, "dur": 76.800,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618058.924, "dur": 5.247,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618061.831, "dur": 0.989,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618070.601, "dur": 1.534,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618079.515, "dur": 9.160,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618083.649, "dur": 4.796,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618168.891, "dur": 190.057,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618171.563, "dur": 2.171,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618177.084, "dur": 181.377,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618178.264, "dur": 0.346,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618179.710, "dur": 22.105,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618203.350, "dur": 5.551,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618205.698, "dur": 2.845,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618209.908, "dur": 26.727,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618213.235, "dur": 1.923,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618216.238, "dur": 20.026,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618219.085, "dur": 2.596,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618238.036, "dur": 21.583,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618261.012, "dur": 13.668,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618276.859, "dur": 14.512,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618292.589, "dur": 12.256,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618306.520, "dur": 23.794,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618308.929, "dur": 1.126,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618314.490, "dur": 0.787,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618331.633, "dur": 12.784,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618345.420, "dur": 12.283,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618364.570, "dur": 1.611,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618373.889, "dur": 3.238,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618375.621, "dur": 0.735,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618434.196, "dur": 51.497,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618490.467, "dur": 9.448,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618495.696, "dur": 3.083,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618501.326, "dur": 24.478,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618547.637, "dur": 6.124,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618549.206, "dur": 3.786,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618551.047, "dur": 1.598,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618556.720, "dur": 46.343,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618560.261, "dur": 42.115,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618606.767, "dur": 16.868,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618628.049, "dur": 25.336,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618630.391, "dur": 22.634,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618635.826, "dur": 1.328,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618658.630, "dur": 26.782,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618660.755, "dur": 24.464,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618665.227, "dur": 4.161,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618670.808, "dur": 14.012,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618699.707, "dur": 7.365,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618704.031, "dur": 2.738,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618708.125, "dur": 1.679,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618709.043, "dur": 0.459,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618751.087, "dur": 22.719,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618776.042, "dur": 14.572,
    "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": 1336753, "tid": 1336753, "ts": 1514656618776.042,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336753, "tid": 1336753,
    "ts": 1514656618796.638, "dur": 6.064,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618800.240, "dur": 1.218,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618807.593, "dur": 6.566,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618811.803, "dur": 1.144,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618815.497, "dur": 3.318,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618817.230, "dur": 1.014,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618822.597, "dur": 8.397,
    "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": 1336753, "tid": 1336753, "ts": 1514656618822.597,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1336753,
    "ts": 1514656618828.410, "dur": 1.328,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618832.141, "dur": 5.777,
    "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": 1336753, "tid": 1336753, "ts": 1514656618832.141,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1336753,
    "ts": 1514656618836.252, "dur": 0.903,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618839.010, "dur": 8.140,
    "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": 1336753, "tid": 1336753, "ts": 1514656618839.010,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1336753,
    "ts": 1514656618845.121, "dur": 0.943,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618848.274, "dur": 4.655,
    "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": 1336753, "tid": 1336753, "ts": 1514656618848.274,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1336753,
    "ts": 1514656618851.245, "dur": 0.959,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618857.185, "dur": 31.149,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618858.798, "dur": 29.334,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618861.299, "dur": 7.315,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618864.215, "dur": 3.844,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618869.373, "dur": 18.166,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618913.372, "dur": 4.016,
    "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": 1336753, "tid": 1336753, "ts": 1514656618913.372,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656618919.861, "dur": 3.022,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618955.641, "dur": 44022.680,
    "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": 1336753, "tid": 1336753, "ts": 1514656618955.641,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 1336753, "tid": 1336753,
    "ts": 1514656618969.967, "dur": 66.568,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618970.918, "dur": 65.426,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618972.166, "dur": 5.477,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618973.608, "dur": 3.613,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656618978.678, "dur": 56.917,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656619054.328, "dur": 26.806,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656619055.451, "dur": 7.088,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656619057.641, "dur": 4.611,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656619064.000, "dur": 16.900,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656619065.636, "dur": 14.844,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656619084.655, "dur": 22.854,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656619085.557, "dur": 6.867,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656619089.328, "dur": 2.868,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656619093.387, "dur": 13.924,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656619094.257, "dur": 12.721,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656619113.802, "dur": 19.659,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656619115.950, "dur": 5.105,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656619121.951, "dur": 11.234,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656619122.466, "dur": 10.449,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656619140.090, "dur": 24.185,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656619166.849, "dur": 53.666,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656619169.179, "dur": 50.914,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656619174.342, "dur": 1.506,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656619177.108, "dur": 25.703,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656619178.634, "dur": 23.965,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656619183.484, "dur": 3.726,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656619187.888, "dur": 14.419,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656619224.768, "dur": 37651.489,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656619226.397, "dur": 37648.146,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656656889.146, "dur": 10.758,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656656896.391, "dur": 1.498,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656656905.250, "dur": 139.881,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656656907.176, "dur": 7.162,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656656909.901, "dur": 3.625,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656656912.141, "dur": 1.088,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656656915.833, "dur": 128.516,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656656920.010, "dur": 122.908,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657050.580, "dur": 5.317,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657053.289, "dur": 0.899,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657064.420, "dur": 3.176,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657077.519, "dur": 8.218,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657080.309, "dur": 5.126,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657222.567, "dur": 225.453,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657225.886, "dur": 2.369,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657230.226, "dur": 217.262,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657235.382, "dur": 0.416,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657237.114, "dur": 34.877,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657273.978, "dur": 4.297,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657276.551, "dur": 1.264,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657279.534, "dur": 29.853,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657283.626, "dur": 1.563,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657289.023, "dur": 19.882,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657292.936, "dur": 3.054,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657311.454, "dur": 24.469,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657337.874, "dur": 15.483,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657357.162, "dur": 16.741,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657375.419, "dur": 14.063,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657391.745, "dur": 25.421,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657394.028, "dur": 1.945,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657400.618, "dur": 0.967,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657418.726, "dur": 13.603,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657433.845, "dur": 12.463,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657454.602, "dur": 1.712,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657463.092, "dur": 4.113,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657465.272, "dur": 0.686,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657557.285, "dur": 74.704,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657638.295, "dur": 9.146,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657643.468, "dur": 1.398,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657648.944, "dur": 29.385,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657684.331, "dur": 8.910,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657686.280, "dur": 6.172,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657688.679, "dur": 3.440,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657696.432, "dur": 47.462,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657700.209, "dur": 43.120,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657749.063, "dur": 17.840,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657773.388, "dur": 3.848,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657775.602, "dur": 0.714,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657781.461, "dur": 53.630,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657782.519, "dur": 6.932,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657783.717, "dur": 5.085,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657787.846, "dur": 0.815,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657790.334, "dur": 44.379,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657791.048, "dur": 43.201,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657838.718, "dur": 3.522,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657840.701, "dur": 0.511,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657848.708, "dur": 1.686,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657860.855, "dur": 7.383,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657863.160, "dur": 4.827,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657965.097, "dur": 237.196,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657967.680, "dur": 2.582,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657974.124, "dur": 227.700,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657975.279, "dur": 0.330,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656657978.938, "dur": 60.079,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658041.629, "dur": 6.415,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658044.229, "dur": 3.300,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658049.112, "dur": 24.722,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658050.232, "dur": 2.313,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658053.845, "dur": 19.656,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658056.857, "dur": 3.177,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658075.380, "dur": 22.832,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658099.512, "dur": 15.773,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658118.499, "dur": 15.271,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658137.584, "dur": 12.745,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658152.394, "dur": 20.815,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658154.414, "dur": 1.259,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658158.008, "dur": 0.840,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658174.852, "dur": 13.237,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658189.356, "dur": 11.253,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658209.540, "dur": 2.460,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658223.249, "dur": 4.069,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658225.973, "dur": 0.447,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658302.702, "dur": 56.109,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658363.768, "dur": 6.933,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658366.381, "dur": 3.045,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658372.164, "dur": 25.083,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658402.193, "dur": 7.644,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658403.600, "dur": 5.564,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658407.731, "dur": 1.262,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658412.374, "dur": 41.108,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658413.659, "dur": 39.182,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658457.861, "dur": 15.790,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658479.032, "dur": 3.694,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658481.147, "dur": 0.727,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658489.013, "dur": 64.317,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658489.992, "dur": 3.968,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658490.787, "dur": 2.571,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658492.486, "dur": 0.641,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658494.600, "dur": 58.243,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658495.557, "dur": 56.330,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658576.936, "dur": 5.663,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658580.189, "dur": 1.160,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658591.511, "dur": 1.877,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658601.905, "dur": 8.812,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658603.861, "dur": 6.512,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658704.055, "dur": 198.186,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658706.276, "dur": 2.459,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658710.067, "dur": 191.578,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658714.231, "dur": 0.346,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658716.164, "dur": 26.050,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658743.953, "dur": 4.007,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658746.422, "dur": 1.272,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658749.533, "dur": 24.051,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658750.439, "dur": 1.608,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658753.361, "dur": 19.919,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658758.390, "dur": 2.691,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658775.114, "dur": 22.650,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658798.900, "dur": 13.848,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658815.377, "dur": 14.931,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658833.910, "dur": 13.062,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658848.878, "dur": 21.423,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658850.750, "dur": 1.728,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658854.945, "dur": 1.092,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658871.598, "dur": 12.070,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658884.932, "dur": 13.205,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658908.409, "dur": 1.431,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658918.019, "dur": 3.513,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656658920.349, "dur": 0.433,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659031.546, "dur": 74.077,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659111.006, "dur": 5.717,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659113.800, "dur": 1.449,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659118.178, "dur": 26.488,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659149.889, "dur": 7.673,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659153.587, "dur": 3.375,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659155.692, "dur": 1.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659164.247, "dur": 45.590,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659165.343, "dur": 43.815,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659214.054, "dur": 16.837,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659237.203, "dur": 5.996,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659241.330, "dur": 1.060,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659247.290, "dur": 50.920,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659248.324, "dur": 6.728,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659249.459, "dur": 4.957,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659251.141, "dur": 3.094,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659255.662, "dur": 42.167,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659256.549, "dur": 40.621,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659304.283, "dur": 3.624,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659306.022, "dur": 0.910,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659313.355, "dur": 1.277,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659322.062, "dur": 6.413,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659324.313, "dur": 3.897,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659417.374, "dur": 208.768,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659420.053, "dur": 2.526,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659423.695, "dur": 201.864,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659424.935, "dur": 0.317,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659429.449, "dur": 20.988,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659451.829, "dur": 3.750,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659454.179, "dur": 1.093,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659456.568, "dur": 26.587,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659457.463, "dur": 3.769,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659464.872, "dur": 18.006,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659467.708, "dur": 2.460,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659484.696, "dur": 20.678,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659506.452, "dur": 14.333,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659523.762, "dur": 32.360,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659558.093, "dur": 13.980,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659574.029, "dur": 22.227,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659575.787, "dur": 1.751,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659579.862, "dur": 1.047,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659600.063, "dur": 12.772,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659614.214, "dur": 10.293,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659633.707, "dur": 2.082,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659644.027, "dur": 5.029,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659647.822, "dur": 0.469,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659716.937, "dur": 54.238,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659775.647, "dur": 8.523,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659781.702, "dur": 1.426,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659785.546, "dur": 23.837,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659813.999, "dur": 7.973,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659815.480, "dur": 5.779,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659817.715, "dur": 3.317,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659824.681, "dur": 42.554,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659825.638, "dur": 40.909,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659873.712, "dur": 15.378,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659894.572, "dur": 3.572,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659896.545, "dur": 0.832,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659901.865, "dur": 49.675,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659902.807, "dur": 6.323,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659903.731, "dur": 4.843,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659907.504, "dur": 0.903,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659909.670, "dur": 41.488,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659910.401, "dur": 40.068,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656659979.778, "dur": 42.205,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660018.673, "dur": 1.048,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660029.858, "dur": 2.006,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660040.578, "dur": 12.180,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660044.842, "dur": 7.675,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660148.758, "dur": 200.633,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660151.137, "dur": 2.312,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660155.005, "dur": 193.903,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660156.925, "dur": 0.385,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660158.460, "dur": 27.999,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660188.081, "dur": 3.415,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660190.334, "dur": 0.842,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660192.525, "dur": 28.269,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660195.841, "dur": 1.715,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660198.803, "dur": 21.699,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660204.425, "dur": 3.109,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660222.448, "dur": 24.038,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660247.910, "dur": 13.729,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660264.986, "dur": 15.670,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660281.945, "dur": 13.540,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660297.549, "dur": 23.119,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660299.114, "dur": 1.553,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660305.326, "dur": 1.084,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660321.839, "dur": 13.270,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660336.382, "dur": 11.725,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660355.684, "dur": 1.544,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660365.868, "dur": 3.650,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660368.320, "dur": 0.454,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660432.351, "dur": 57.691,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660494.648, "dur": 7.012,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660499.572, "dur": 1.038,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660503.020, "dur": 40.046,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660550.676, "dur": 6.924,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660552.444, "dur": 4.329,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660554.561, "dur": 1.969,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660560.848, "dur": 47.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660564.513, "dur": 42.600,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660612.029, "dur": 17.056,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660635.444, "dur": 3.957,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660637.509, "dur": 0.976,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660643.287, "dur": 87.530,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660676.402, "dur": 6.853,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660679.721, "dur": 2.833,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660681.469, "dur": 0.860,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660683.901, "dur": 46.478,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660684.452, "dur": 45.332,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660734.842, "dur": 5.863,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660736.689, "dur": 2.934,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660749.715, "dur": 1.452,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660759.386, "dur": 6.358,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660761.387, "dur": 4.106,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660895.630, "dur": 248.449,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660898.792, "dur": 3.367,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660903.846, "dur": 239.590,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660907.521, "dur": 0.414,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660908.988, "dur": 27.052,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660937.577, "dur": 3.555,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660939.916, "dur": 0.898,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660942.168, "dur": 25.742,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660943.186, "dur": 1.887,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660946.485, "dur": 21.071,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660951.494, "dur": 2.552,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656660969.576, "dur": 63.012,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661034.968, "dur": 16.272,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661054.414, "dur": 15.469,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661073.714, "dur": 13.912,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661089.912, "dur": 22.730,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661091.834, "dur": 1.879,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661096.166, "dur": 1.267,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661114.033, "dur": 12.905,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661128.284, "dur": 11.523,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661152.983, "dur": 2.306,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661165.575, "dur": 4.016,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661168.255, "dur": 0.502,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661243.038, "dur": 57.224,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661304.708, "dur": 4.894,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661307.472, "dur": 1.110,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661311.285, "dur": 24.934,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661340.550, "dur": 8.354,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661344.485, "dur": 3.569,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661346.555, "dur": 1.273,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661351.734, "dur": 40.583,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661353.023, "dur": 38.648,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661396.042, "dur": 14.617,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661416.149, "dur": 5.764,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661420.287, "dur": 0.806,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661425.680, "dur": 47.987,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661426.720, "dur": 5.203,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661427.598, "dur": 3.703,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661428.648, "dur": 2.512,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661432.518, "dur": 40.528,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661433.142, "dur": 39.018,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661477.360, "dur": 3.888,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661479.473, "dur": 0.828,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661488.517, "dur": 1.374,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661497.121, "dur": 5.437,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661498.988, "dur": 3.256,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661602.592, "dur": 285.845,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661604.982, "dur": 2.789,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661612.007, "dur": 275.750,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661613.374, "dur": 0.340,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661700.102, "dur": 27.289,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661729.267, "dur": 3.173,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661731.199, "dur": 0.952,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661733.465, "dur": 24.921,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661734.835, "dur": 3.730,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661739.876, "dur": 18.150,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661742.479, "dur": 2.606,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661762.587, "dur": 22.538,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661786.391, "dur": 13.844,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661803.200, "dur": 15.982,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661820.673, "dur": 13.044,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661835.661, "dur": 20.665,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661837.157, "dur": 1.283,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661840.809, "dur": 1.152,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661857.676, "dur": 12.736,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661874.284, "dur": 12.448,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661895.761, "dur": 2.225,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661906.648, "dur": 3.666,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661909.192, "dur": 0.383,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656661975.102, "dur": 108.351,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662090.399, "dur": 5.498,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662093.483, "dur": 1.067,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662097.298, "dur": 26.858,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662131.940, "dur": 6.003,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662133.741, "dur": 3.366,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662135.534, "dur": 1.370,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662141.136, "dur": 41.285,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662142.408, "dur": 39.401,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662186.762, "dur": 16.315,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662209.231, "dur": 6.128,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662213.508, "dur": 1.026,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662219.206, "dur": 55.478,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662220.388, "dur": 6.360,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662221.134, "dur": 4.834,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662224.957, "dur": 0.899,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662227.309, "dur": 46.961,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662230.274, "dur": 43.103,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662278.350, "dur": 3.623,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662279.952, "dur": 0.866,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662287.982, "dur": 1.862,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662297.249, "dur": 8.500,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662301.652, "dur": 3.781,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662389.074, "dur": 218.957,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662391.482, "dur": 2.725,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662395.397, "dur": 212.095,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662398.675, "dur": 0.290,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662400.022, "dur": 21.237,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662423.316, "dur": 4.054,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662425.653, "dur": 1.266,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662428.267, "dur": 32.065,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662438.996, "dur": 1.733,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662442.006, "dur": 18.092,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662444.678, "dur": 2.700,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662461.738, "dur": 21.028,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662484.342, "dur": 15.962,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662503.193, "dur": 15.268,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662519.617, "dur": 29.925,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662552.393, "dur": 26.950,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662554.690, "dur": 2.110,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662561.740, "dur": 0.850,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662580.654, "dur": 13.404,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662595.365, "dur": 11.119,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662615.211, "dur": 2.399,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662626.747, "dur": 3.663,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662629.361, "dur": 0.292,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662696.816, "dur": 52.059,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662753.566, "dur": 9.492,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662758.822, "dur": 3.080,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662764.211, "dur": 25.069,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662794.394, "dur": 5.341,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662795.807, "dur": 3.341,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662797.956, "dur": 1.013,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662802.279, "dur": 44.569,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662805.652, "dur": 40.436,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662850.953, "dur": 15.159,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662870.746, "dur": 25.709,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662873.272, "dur": 22.777,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662878.845, "dur": 1.060,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662901.059, "dur": 29.832,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662903.376, "dur": 27.308,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662907.818, "dur": 4.268,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662916.049, "dur": 14.131,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662944.609, "dur": 5.139,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662946.869, "dur": 2.419,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662950.880, "dur": 1.162,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656662951.347, "dur": 0.625,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663033.744, "dur": 27.308,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663062.985, "dur": 16.182,
    "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": 1336753, "tid": 1336753, "ts": 1514656663062.985,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336753, "tid": 1336753,
    "ts": 1514656663085.520, "dur": 7.091,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663089.847, "dur": 1.171,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663110.023, "dur": 9.248,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663117.086, "dur": 0.940,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663120.633, "dur": 3.236,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663122.396, "dur": 0.891,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663131.069, "dur": 6.137,
    "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": 1336753, "tid": 1336753, "ts": 1514656663131.069,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1336753,
    "ts": 1514656663134.493, "dur": 1.399,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663138.217, "dur": 6.095,
    "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": 1336753, "tid": 1336753, "ts": 1514656663138.217,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1336753,
    "ts": 1514656663142.642, "dur": 0.827,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663145.411, "dur": 7.625,
    "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": 1336753, "tid": 1336753, "ts": 1514656663145.411,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1336753,
    "ts": 1514656663151.138, "dur": 1.029,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663154.091, "dur": 6.465,
    "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": 1336753, "tid": 1336753, "ts": 1514656663154.091,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336753, "tid": 1336753,
    "ts": 1514656663156.918, "dur": 2.815,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663164.732, "dur": 32.822,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663166.250, "dur": 31.102,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663169.096, "dur": 7.189,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663171.650, "dur": 4.126,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663177.517, "dur": 19.316,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663223.974, "dur": 6.447,
    "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": 1336753, "tid": 1336753, "ts": 1514656663223.974,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336753, "tid": 1336753,
    "ts": 1514656663232.711, "dur": 1.323,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663265.780, "dur": 43598.430,
    "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": 1336753, "tid": 1336753, "ts": 1514656663265.780,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 1336753, "tid": 1336753,
    "ts": 1514656663279.784, "dur": 27.796,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663280.779, "dur": 26.581,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663282.198, "dur": 7.072,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663283.719, "dur": 5.175,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663289.878, "dur": 17.050,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663323.028, "dur": 28.900,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663324.553, "dur": 8.903,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663328.675, "dur": 4.479,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663334.869, "dur": 16.869,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663336.935, "dur": 14.447,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663355.371, "dur": 20.981,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663356.179, "dur": 4.864,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663357.914, "dur": 2.889,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663361.521, "dur": 14.643,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663362.345, "dur": 13.465,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663381.489, "dur": 21.623,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663385.653, "dur": 3.646,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663390.227, "dur": 12.634,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663390.885, "dur": 11.689,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663407.703, "dur": 22.885,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663433.196, "dur": 50.439,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663435.283, "dur": 47.985,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663442.295, "dur": 0.815,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663444.347, "dur": 22.434,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663446.010, "dur": 20.538,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663448.378, "dur": 2.893,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663452.184, "dur": 14.013,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663489.435, "dur": 37534.188,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656663491.499, "dur": 37530.337,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701038.376, "dur": 9.266,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701044.153, "dur": 1.401,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701053.119, "dur": 109.463,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701055.264, "dur": 9.406,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701057.634, "dur": 6.131,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701062.280, "dur": 1.226,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701066.132, "dur": 95.629,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701068.005, "dur": 92.657,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701166.211, "dur": 4.161,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701168.222, "dur": 0.690,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701178.357, "dur": 2.978,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701193.692, "dur": 7.748,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701196.403, "dur": 4.777,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701335.511, "dur": 238.272,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701341.266, "dur": 2.526,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701345.480, "dur": 227.826,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701347.384, "dur": 0.411,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701349.415, "dur": 28.878,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701382.832, "dur": 5.981,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701387.554, "dur": 0.792,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701390.139, "dur": 24.501,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701391.215, "dur": 1.485,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701393.976, "dur": 20.352,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701398.376, "dur": 3.222,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701416.415, "dur": 24.752,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701443.213, "dur": 18.579,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701464.759, "dur": 16.669,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701482.873, "dur": 15.302,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701502.683, "dur": 22.378,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701505.023, "dur": 1.600,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701508.935, "dur": 1.115,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701526.627, "dur": 31.063,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701559.860, "dur": 11.833,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701582.790, "dur": 2.817,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701595.257, "dur": 4.997,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701597.977, "dur": 1.101,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701677.879, "dur": 68.119,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701751.829, "dur": 7.016,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701754.899, "dur": 1.350,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701760.435, "dur": 27.281,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701792.990, "dur": 9.339,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701794.959, "dur": 6.714,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701799.639, "dur": 1.803,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701805.293, "dur": 42.964,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701806.760, "dur": 40.659,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701852.567, "dur": 16.467,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701875.264, "dur": 4.068,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701877.572, "dur": 0.873,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701886.004, "dur": 49.148,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701887.354, "dur": 3.877,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701887.998, "dur": 2.640,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701889.490, "dur": 0.928,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701892.082, "dur": 42.667,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701893.118, "dur": 40.990,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701941.236, "dur": 4.226,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701943.352, "dur": 0.993,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701951.238, "dur": 1.377,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701960.874, "dur": 8.807,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656701962.777, "dur": 6.583,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702107.292, "dur": 204.115,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702109.559, "dur": 3.658,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702115.431, "dur": 195.529,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702117.264, "dur": 0.706,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702119.093, "dur": 26.838,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702147.671, "dur": 3.676,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702150.110, "dur": 0.961,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702152.510, "dur": 28.882,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702153.902, "dur": 1.516,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702156.985, "dur": 24.110,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702164.223, "dur": 3.097,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702183.046, "dur": 23.629,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702208.590, "dur": 14.438,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702225.661, "dur": 15.770,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702242.887, "dur": 13.762,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702258.401, "dur": 21.804,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702260.582, "dur": 1.337,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702263.857, "dur": 0.922,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702283.745, "dur": 13.165,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702298.069, "dur": 12.129,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702318.034, "dur": 2.405,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702329.339, "dur": 3.670,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702331.495, "dur": 0.594,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702400.430, "dur": 54.054,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702459.878, "dur": 4.781,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702462.132, "dur": 1.393,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702468.396, "dur": 25.518,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702498.264, "dur": 6.524,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702500.177, "dur": 4.000,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702502.391, "dur": 1.567,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702507.153, "dur": 57.112,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702508.517, "dur": 54.545,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702569.509, "dur": 17.056,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702595.596, "dur": 4.669,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702598.209, "dur": 1.027,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702604.242, "dur": 56.243,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702605.461, "dur": 6.067,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702606.677, "dur": 4.321,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702608.045, "dur": 2.772,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702614.274, "dur": 45.828,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702615.392, "dur": 44.050,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702664.178, "dur": 4.318,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702666.489, "dur": 0.736,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702673.783, "dur": 1.305,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702683.373, "dur": 6.313,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702685.423, "dur": 4.040,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702780.405, "dur": 187.380,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702782.748, "dur": 2.305,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702786.573, "dur": 180.725,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702787.706, "dur": 0.308,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702791.314, "dur": 21.677,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702814.591, "dur": 4.093,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702817.058, "dur": 1.314,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702819.664, "dur": 26.183,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702820.816, "dur": 3.036,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702825.118, "dur": 20.182,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702829.833, "dur": 2.551,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702847.067, "dur": 20.431,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702868.893, "dur": 14.463,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702885.776, "dur": 14.374,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702901.446, "dur": 13.871,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702916.887, "dur": 20.790,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702918.964, "dur": 1.799,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702923.016, "dur": 0.614,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702941.408, "dur": 12.308,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702955.105, "dur": 11.034,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656702973.962, "dur": 1.670,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703019.669, "dur": 4.669,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703022.174, "dur": 0.888,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703097.869, "dur": 81.924,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703184.941, "dur": 7.588,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703190.311, "dur": 1.015,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703193.915, "dur": 34.966,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703233.260, "dur": 7.127,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703234.705, "dur": 5.046,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703236.618, "dur": 2.914,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703242.807, "dur": 53.927,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703244.075, "dur": 51.919,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703302.300, "dur": 17.334,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703325.040, "dur": 3.549,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703327.059, "dur": 0.756,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703331.929, "dur": 50.225,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703332.972, "dur": 5.699,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703333.532, "dur": 4.662,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703337.069, "dur": 1.016,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703339.328, "dur": 42.478,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703340.321, "dur": 40.831,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703385.862, "dur": 16.767,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703400.754, "dur": 0.628,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703407.931, "dur": 1.553,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703419.187, "dur": 8.548,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703421.007, "dur": 6.455,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703509.899, "dur": 215.443,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703512.106, "dur": 2.019,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703515.695, "dur": 209.247,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703517.042, "dur": 0.513,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703518.554, "dur": 42.963,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703566.603, "dur": 3.489,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703568.449, "dur": 1.331,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703571.144, "dur": 27.001,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703572.481, "dur": 2.444,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703576.148, "dur": 21.619,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703582.162, "dur": 2.826,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703599.557, "dur": 22.500,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703623.628, "dur": 13.911,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703640.219, "dur": 15.319,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703656.981, "dur": 14.320,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703675.444, "dur": 21.110,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703677.717, "dur": 1.513,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703681.511, "dur": 0.767,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703697.725, "dur": 12.891,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703711.702, "dur": 12.295,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703733.642, "dur": 2.238,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703744.369, "dur": 3.508,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703746.567, "dur": 0.425,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703816.976, "dur": 59.440,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703881.112, "dur": 5.322,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703883.852, "dur": 1.421,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703887.806, "dur": 25.239,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703917.499, "dur": 7.571,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703919.073, "dur": 5.386,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703923.211, "dur": 1.059,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703927.730, "dur": 42.742,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703928.951, "dur": 40.837,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656703973.987, "dur": 53.577,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704036.001, "dur": 4.451,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704038.548, "dur": 0.755,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704044.680, "dur": 59.844,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704047.979, "dur": 6.277,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704049.256, "dur": 4.367,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704050.623, "dur": 2.844,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704055.298, "dur": 48.812,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704056.366, "dur": 47.039,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704108.426, "dur": 4.295,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704110.799, "dur": 0.699,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704118.775, "dur": 1.503,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704130.969, "dur": 6.346,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704132.654, "dur": 4.440,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704222.923, "dur": 199.247,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704225.479, "dur": 1.964,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704229.436, "dur": 192.285,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704230.796, "dur": 0.254,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704236.529, "dur": 23.201,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704266.993, "dur": 3.976,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704269.728, "dur": 0.905,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704271.941, "dur": 25.729,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704273.468, "dur": 3.459,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704278.399, "dur": 18.958,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704281.154, "dur": 2.931,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704298.950, "dur": 20.760,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704321.210, "dur": 15.077,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704340.749, "dur": 14.384,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704356.431, "dur": 13.657,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704371.857, "dur": 22.191,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704373.829, "dur": 1.785,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704377.688, "dur": 1.064,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704395.135, "dur": 13.074,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704409.541, "dur": 11.411,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704431.326, "dur": 1.626,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704440.791, "dur": 3.325,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704443.000, "dur": 0.312,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704505.541, "dur": 66.851,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704578.500, "dur": 5.951,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704581.459, "dur": 1.586,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704585.693, "dur": 25.275,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704616.531, "dur": 9.838,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704619.978, "dur": 5.748,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704621.899, "dur": 3.645,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704629.473, "dur": 42.899,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704630.477, "dur": 41.381,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704676.363, "dur": 15.729,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704697.772, "dur": 6.199,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704702.256, "dur": 0.864,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704708.018, "dur": 48.059,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704709.099, "dur": 3.914,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704709.865, "dur": 2.621,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704711.303, "dur": 1.054,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704713.661, "dur": 42.009,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704714.476, "dur": 40.577,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704762.151, "dur": 4.341,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704764.613, "dur": 0.684,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704772.062, "dur": 1.738,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704781.633, "dur": 6.118,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704783.086, "dur": 4.447,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704870.335, "dur": 229.540,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704872.434, "dur": 2.937,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704878.862, "dur": 220.522,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704879.986, "dur": 0.247,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704881.075, "dur": 20.945,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704903.612, "dur": 4.864,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704905.438, "dur": 2.718,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704909.413, "dur": 23.459,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704910.518, "dur": 1.521,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704913.199, "dur": 19.373,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704917.734, "dur": 2.366,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704934.264, "dur": 21.648,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704957.235, "dur": 13.475,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656704973.493, "dur": 50.679,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705026.562, "dur": 14.614,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705042.912, "dur": 25.119,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705044.916, "dur": 2.601,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705049.638, "dur": 3.047,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705071.864, "dur": 13.155,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705086.386, "dur": 11.959,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705107.537, "dur": 2.416,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705118.821, "dur": 3.543,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705120.993, "dur": 0.532,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705185.607, "dur": 53.709,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705243.588, "dur": 6.512,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705248.184, "dur": 0.907,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705251.434, "dur": 25.710,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705281.305, "dur": 5.746,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705282.446, "dur": 3.952,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705284.441, "dur": 1.687,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705289.904, "dur": 41.398,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705291.238, "dur": 39.485,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705337.127, "dur": 15.576,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705358.261, "dur": 3.940,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705360.563, "dur": 0.758,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705365.722, "dur": 48.925,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705366.790, "dur": 5.500,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705367.640, "dur": 4.156,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705370.996, "dur": 0.669,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705372.997, "dur": 41.222,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705373.881, "dur": 39.833,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705418.005, "dur": 3.985,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705420.090, "dur": 0.838,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705426.994, "dur": 1.579,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705435.752, "dur": 7.928,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705439.547, "dur": 3.880,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705519.186, "dur": 209.779,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705523.353, "dur": 2.300,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705541.141, "dur": 187.107,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705543.505, "dur": 0.308,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705545.165, "dur": 25.152,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705571.970, "dur": 5.688,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705576.184, "dur": 1.099,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705578.824, "dur": 25.988,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705582.276, "dur": 1.603,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705585.123, "dur": 19.436,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705588.019, "dur": 2.475,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705606.113, "dur": 20.304,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705627.951, "dur": 17.068,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705647.574, "dur": 14.020,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705663.008, "dur": 13.231,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705677.705, "dur": 23.051,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705679.532, "dur": 1.322,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705685.435, "dur": 0.830,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705701.964, "dur": 12.878,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705715.964, "dur": 11.469,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705736.455, "dur": 2.180,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705747.032, "dur": 3.392,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705749.125, "dur": 0.609,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705811.404, "dur": 49.492,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705864.931, "dur": 7.040,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705869.995, "dur": 0.893,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705873.187, "dur": 22.701,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705899.826, "dur": 4.611,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705901.040, "dur": 2.802,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705902.855, "dur": 0.853,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705906.974, "dur": 41.958,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705910.327, "dur": 37.945,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705952.412, "dur": 15.310,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705972.959, "dur": 3.624,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705975.106, "dur": 0.672,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705980.319, "dur": 92.873,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705981.443, "dur": 42.044,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656705982.724, "dur": 39.857,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706021.179, "dur": 1.013,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706024.612, "dur": 48.078,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706025.929, "dur": 46.211,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706078.095, "dur": 4.752,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706080.585, "dur": 0.960,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706089.149, "dur": 1.536,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706100.351, "dur": 7.960,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706102.201, "dur": 5.846,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706193.059, "dur": 194.533,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706195.289, "dur": 2.173,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706199.163, "dur": 187.997,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706200.659, "dur": 0.316,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706204.233, "dur": 25.940,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706231.665, "dur": 2.910,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706233.526, "dur": 0.652,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706235.902, "dur": 24.933,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706237.271, "dur": 1.930,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706240.554, "dur": 20.026,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706245.662, "dur": 2.395,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706262.333, "dur": 21.582,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706285.697, "dur": 13.400,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706301.300, "dur": 16.381,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706321.217, "dur": 12.861,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706335.602, "dur": 21.110,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706337.810, "dur": 1.471,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706341.034, "dur": 0.965,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706357.870, "dur": 12.592,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706371.557, "dur": 14.839,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706393.953, "dur": 1.677,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706405.454, "dur": 3.340,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706407.469, "dur": 0.472,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706470.800, "dur": 48.447,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706523.680, "dur": 22.096,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706526.646, "dur": 17.120,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706547.738, "dur": 27.074,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706579.909, "dur": 7.676,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706581.180, "dur": 5.778,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706585.727, "dur": 1.046,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706590.278, "dur": 44.701,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706591.640, "dur": 42.631,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706638.736, "dur": 18.611,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706662.155, "dur": 26.774,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706664.797, "dur": 23.732,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706672.006, "dur": 0.886,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706693.873, "dur": 122.296,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706695.750, "dur": 120.127,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706793.131, "dur": 5.448,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706800.074, "dur": 15.111,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706831.761, "dur": 4.668,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706833.515, "dur": 2.649,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706837.425, "dur": 1.023,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706837.795, "dur": 0.592,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706882.591, "dur": 24.932,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656706909.366, "dur": 13.878,
    "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": 1336753, "tid": 1336753, "ts": 1514656706909.366,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward", "pid": 1336753, "tid": 1336753,
    "ts": 1514656707070.686, "dur": 40.790,
    "args": {
      "External id": 2482697,"Record function id": 0, "Ev Idx": 10400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::div", "pid": 1336753, "tid": 1336753,
    "ts": 1514656707218.117, "dur": 38.618,
    "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": 1336753, "tid": 1336753, "ts": 1514656707218.117,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::ones_like", "pid": 1336753, "tid": 1336753,
    "ts": 1514656707316.080, "dur": 28.084,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656707317.618, "dur": 8.733,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656707321.478, "dur": 4.252,
    "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": 1336753, "tid": 1336753,
    "ts": 1514656707327.806, "dur": 16.044,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658027157.499, "dur": 79.448,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658027249.294, "dur": 24.367,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658027282.054, "dur": 24.750,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658027899.766, "dur": 31.499,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658028449.976, "dur": 37.086,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658029955.271, "dur": 3178.733,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658030612.619, "dur": 974.782,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658030634.432, "dur": 65.034,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658030637.692, "dur": 11.300,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658030652.133, "dur": 47.011,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658030654.368, "dur": 44.069,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033183.068, "dur": 4.378,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033190.063, "dur": 0.517,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033191.939, "dur": 0.336,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033193.492, "dur": 0.241,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033195.207, "dur": 0.629,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033196.977, "dur": 0.304,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033198.587, "dur": 0.208,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033202.462, "dur": 0.304,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033203.849, "dur": 0.219,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033205.130, "dur": 0.353,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033206.541, "dur": 0.215,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033207.902, "dur": 0.234,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033208.895, "dur": 0.341,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033210.172, "dur": 0.220,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033211.225, "dur": 0.205,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033214.788, "dur": 0.218,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033215.799, "dur": 0.206,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033217.091, "dur": 0.219,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033218.030, "dur": 0.303,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033219.052, "dur": 0.207,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033220.181, "dur": 0.237,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033221.310, "dur": 0.357,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033222.676, "dur": 0.213,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033226.490, "dur": 0.342,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033227.645, "dur": 0.318,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033228.955, "dur": 0.404,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033230.302, "dur": 0.212,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033231.647, "dur": 0.217,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033232.841, "dur": 0.252,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033234.059, "dur": 0.213,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033235.128, "dur": 0.203,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033238.374, "dur": 0.209,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033239.515, "dur": 0.218,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033240.852, "dur": 0.201,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033242.166, "dur": 0.374,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033243.281, "dur": 0.205,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033244.200, "dur": 0.212,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033245.139, "dur": 0.207,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033246.270, "dur": 0.329,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033249.704, "dur": 0.375,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033250.839, "dur": 0.354,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033252.117, "dur": 0.319,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033253.356, "dur": 0.361,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033254.526, "dur": 0.220,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033255.481, "dur": 0.202,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033258.283, "dur": 0.213,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033259.294, "dur": 0.202,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033262.599, "dur": 0.206,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033263.728, "dur": 0.204,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033264.639, "dur": 0.205,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033265.749, "dur": 0.201,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033266.743, "dur": 0.207,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033267.698, "dur": 0.220,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033269.092, "dur": 0.202,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033270.016, "dur": 0.202,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033273.292, "dur": 0.205,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033274.266, "dur": 0.202,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033275.179, "dur": 0.362,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033276.292, "dur": 0.326,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033277.358, "dur": 0.212,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033278.336, "dur": 0.324,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033279.515, "dur": 0.235,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033280.743, "dur": 0.301,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033283.896, "dur": 0.446,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033285.316, "dur": 0.373,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033286.857, "dur": 0.379,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033287.958, "dur": 0.333,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033289.055, "dur": 0.209,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033290.003, "dur": 0.327,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033291.038, "dur": 0.208,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033292.140, "dur": 0.210,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033295.159, "dur": 0.211,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033296.234, "dur": 0.203,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033297.291, "dur": 0.219,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033298.340, "dur": 0.208,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033299.272, "dur": 0.202,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033300.342, "dur": 0.205,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033302.554, "dur": 0.589,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033303.997, "dur": 0.201,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033307.081, "dur": 0.273,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033308.070, "dur": 0.328,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033309.400, "dur": 0.300,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033310.797, "dur": 0.380,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033311.940, "dur": 0.205,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033312.863, "dur": 0.315,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033313.887, "dur": 0.307,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033314.931, "dur": 0.202,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033317.987, "dur": 0.211,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033318.982, "dur": 0.208,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033319.896, "dur": 0.206,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033320.838, "dur": 0.202,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033321.779, "dur": 0.209,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033322.864, "dur": 0.205,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033324.671, "dur": 0.206,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033325.622, "dur": 0.202,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033328.935, "dur": 0.207,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033329.884, "dur": 0.204,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033330.795, "dur": 0.228,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033331.790, "dur": 0.223,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033332.745, "dur": 0.202,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033333.676, "dur": 0.201,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033334.622, "dur": 0.207,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033335.568, "dur": 0.205,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033338.968, "dur": 0.207,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033339.887, "dur": 0.202,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033340.834, "dur": 0.202,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033341.858, "dur": 0.197,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033342.813, "dur": 0.209,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033343.844, "dur": 0.218,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033345.202, "dur": 0.225,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033346.448, "dur": 0.202,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033350.022, "dur": 0.348,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033351.190, "dur": 0.292,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033352.387, "dur": 0.384,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033353.631, "dur": 0.307,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033354.687, "dur": 0.208,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033355.839, "dur": 0.199,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033356.908, "dur": 0.219,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033357.944, "dur": 0.205,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033361.138, "dur": 0.327,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033362.189, "dur": 0.342,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033363.264, "dur": 0.203,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033364.180, "dur": 0.200,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033365.100, "dur": 0.202,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033366.253, "dur": 0.220,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033367.227, "dur": 0.208,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033368.303, "dur": 0.206,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033371.667, "dur": 0.208,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033372.604, "dur": 0.202,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033374.133, "dur": 0.208,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033375.101, "dur": 0.364,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033376.206, "dur": 0.207,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033377.140, "dur": 0.202,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033378.128, "dur": 0.211,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033379.105, "dur": 0.206,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033382.565, "dur": 0.221,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033383.573, "dur": 0.421,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033384.737, "dur": 0.435,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033385.883, "dur": 0.420,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033387.004, "dur": 0.201,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033387.973, "dur": 0.200,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033388.940, "dur": 0.439,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033390.167, "dur": 0.235,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033393.177, "dur": 0.402,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033394.389, "dur": 0.433,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033395.574, "dur": 0.221,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033396.504, "dur": 0.200,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033397.408, "dur": 0.203,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033398.340, "dur": 0.204,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033399.812, "dur": 0.206,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033401.052, "dur": 0.271,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033404.346, "dur": 0.323,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033405.383, "dur": 0.347,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033406.472, "dur": 0.367,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033407.809, "dur": 0.205,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033408.725, "dur": 0.206,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033409.799, "dur": 0.330,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033410.841, "dur": 0.343,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033411.991, "dur": 0.329,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033416.542, "dur": 0.325,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033417.745, "dur": 0.199,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033418.878, "dur": 0.231,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033419.842, "dur": 0.215,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033420.771, "dur": 0.205,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033421.841, "dur": 0.203,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033422.755, "dur": 0.204,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033423.683, "dur": 0.203,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033426.615, "dur": 0.208,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033427.566, "dur": 0.206,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033428.690, "dur": 0.220,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033429.699, "dur": 0.210,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033430.650, "dur": 0.205,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033431.744, "dur": 0.206,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033433.019, "dur": 0.198,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033433.957, "dur": 0.201,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033437.052, "dur": 0.209,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033438.017, "dur": 0.216,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033439.133, "dur": 0.234,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033440.463, "dur": 0.206,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033441.556, "dur": 0.204,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033442.496, "dur": 0.201,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033443.403, "dur": 0.203,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033444.322, "dur": 0.201,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033447.254, "dur": 0.206,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033448.206, "dur": 0.203,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033449.164, "dur": 0.207,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033450.141, "dur": 0.201,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033451.110, "dur": 0.203,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033452.084, "dur": 0.221,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033453.010, "dur": 0.221,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033454.003, "dur": 0.197,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033456.919, "dur": 0.203,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033458.010, "dur": 0.244,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033458.955, "dur": 0.207,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033459.947, "dur": 0.199,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033460.908, "dur": 0.205,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033462.043, "dur": 0.232,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033463.009, "dur": 0.311,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033464.231, "dur": 0.244,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033467.235, "dur": 0.495,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033468.575, "dur": 0.230,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033469.909, "dur": 0.289,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033471.024, "dur": 0.355,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033472.082, "dur": 0.206,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033473.045, "dur": 0.476,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033474.326, "dur": 0.248,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033475.297, "dur": 0.329,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033478.348, "dur": 0.209,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033479.281, "dur": 0.225,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033480.217, "dur": 0.515,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033481.512, "dur": 0.405,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033482.669, "dur": 0.204,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033483.810, "dur": 0.210,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033484.738, "dur": 0.203,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033485.863, "dur": 0.205,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033489.140, "dur": 0.470,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033490.379, "dur": 0.402,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033491.675, "dur": 0.219,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033492.813, "dur": 0.226,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033493.746, "dur": 0.221,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033494.886, "dur": 0.204,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033495.794, "dur": 0.208,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033496.880, "dur": 0.206,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033499.908, "dur": 0.543,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033501.166, "dur": 0.314,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033502.187, "dur": 0.314,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033503.224, "dur": 0.319,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033504.245, "dur": 0.204,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033505.306, "dur": 0.203,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033506.273, "dur": 0.312,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033507.294, "dur": 0.340,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033510.581, "dur": 0.305,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033511.627, "dur": 0.334,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033513.126, "dur": 0.212,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033514.068, "dur": 0.207,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033514.992, "dur": 0.225,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033515.934, "dur": 0.199,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033516.854, "dur": 0.205,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033517.761, "dur": 0.204,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033520.901, "dur": 0.213,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033521.925, "dur": 0.196,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033522.998, "dur": 0.205,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033523.993, "dur": 0.220,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033524.920, "dur": 0.200,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033526.058, "dur": 0.199,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033527.021, "dur": 0.204,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033538.058, "dur": 0.809,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033544.309, "dur": 0.264,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033545.378, "dur": 0.213,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033546.383, "dur": 0.239,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033547.349, "dur": 0.203,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033548.341, "dur": 0.227,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033549.375, "dur": 0.210,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033550.314, "dur": 0.202,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033551.341, "dur": 0.225,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033554.554, "dur": 0.203,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033567.840, "dur": 0.345,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033569.625, "dur": 0.218,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033570.724, "dur": 0.223,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033571.748, "dur": 0.208,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033572.872, "dur": 0.217,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033573.842, "dur": 0.212,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033575.007, "dur": 0.211,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033578.129, "dur": 0.370,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033579.258, "dur": 0.318,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033580.481, "dur": 0.343,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033581.613, "dur": 0.324,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033582.718, "dur": 0.199,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033583.761, "dur": 0.339,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033586.467, "dur": 0.214,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033587.545, "dur": 0.375,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033590.779, "dur": 0.367,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033591.989, "dur": 0.367,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033594.660, "dur": 0.239,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033595.762, "dur": 0.204,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033596.696, "dur": 0.200,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033597.613, "dur": 0.199,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033598.540, "dur": 0.208,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033599.521, "dur": 0.201,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033602.492, "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": 1336753, "tid": 1336753,
    "ts": 1514658033603.432, "dur": 0.213,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033604.372, "dur": 0.202,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033605.348, "dur": 0.481,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033606.628, "dur": 0.228,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033607.841, "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": 1336753, "tid": 1336753,
    "ts": 1514658033608.785, "dur": 0.251,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033609.901, "dur": 0.197,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033612.956, "dur": 0.429,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033614.294, "dur": 0.434,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033615.432, "dur": 0.436,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033616.564, "dur": 0.431,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658033674.683, "dur": 1770.884,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034170.327, "dur": 1199.708,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034179.716, "dur": 8.476,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034184.000, "dur": 3.401,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034188.895, "dur": 3.960,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034189.838, "dur": 2.896,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034193.578, "dur": 3.319,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034195.814, "dur": 0.979,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034197.173, "dur": 2.063,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034197.753, "dur": 1.231,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034201.441, "dur": 5.150,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034205.924, "dur": 0.600,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034206.968, "dur": 1.884,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034207.928, "dur": 0.856,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034209.122, "dur": 1.998,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034209.982, "dur": 1.075,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034211.407, "dur": 3.669,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034214.044, "dur": 0.955,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034215.295, "dur": 1.511,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034216.072, "dur": 0.664,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034217.028, "dur": 3.384,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034217.821, "dur": 2.505,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034220.657, "dur": 3.367,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034222.687, "dur": 1.091,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034224.332, "dur": 1.366,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034224.682, "dur": 0.947,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034228.215, "dur": 4.791,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034232.359, "dur": 0.575,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034233.268, "dur": 1.595,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034233.904, "dur": 0.892,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034235.108, "dur": 1.840,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034235.842, "dur": 1.038,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034237.171, "dur": 3.733,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034240.091, "dur": 0.741,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034241.126, "dur": 1.669,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034241.721, "dur": 0.815,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034243.208, "dur": 3.524,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034244.050, "dur": 2.279,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034246.988, "dur": 3.769,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034249.717, "dur": 0.973,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034251.177, "dur": 1.358,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034251.788, "dur": 0.680,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034255.131, "dur": 5.044,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034259.343, "dur": 0.762,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034260.418, "dur": 1.686,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034260.958, "dur": 0.902,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034262.336, "dur": 1.510,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034262.885, "dur": 0.896,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034264.100, "dur": 4.003,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034267.277, "dur": 0.757,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034268.326, "dur": 2.110,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034269.122, "dur": 0.966,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034270.660, "dur": 3.521,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034271.480, "dur": 2.629,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034274.417, "dur": 3.692,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034277.065, "dur": 0.973,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034278.356, "dur": 1.659,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034278.708, "dur": 1.242,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034282.522, "dur": 5.486,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034287.252, "dur": 0.690,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034288.255, "dur": 1.913,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034288.759, "dur": 1.158,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034290.388, "dur": 1.595,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034291.087, "dur": 0.831,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034292.225, "dur": 6.063,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034297.238, "dur": 0.983,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034298.508, "dur": 1.330,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034299.030, "dur": 0.740,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034300.061, "dur": 3.713,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034300.731, "dur": 2.570,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034303.998, "dur": 5.354,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034308.459, "dur": 0.827,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034309.574, "dur": 2.931,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034310.156, "dur": 2.280,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034314.830, "dur": 3.050,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034317.077, "dur": 0.737,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034318.320, "dur": 1.476,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034318.879, "dur": 0.853,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034320.042, "dur": 3.361,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034322.468, "dur": 0.867,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034323.667, "dur": 2.800,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034325.686, "dur": 0.712,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034326.713, "dur": 1.326,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034327.342, "dur": 0.633,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034328.263, "dur": 3.064,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034330.538, "dur": 0.720,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034331.553, "dur": 3.057,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034333.526, "dur": 0.805,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034334.837, "dur": 3.075,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034335.351, "dur": 2.491,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034340.100, "dur": 2.659,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034341.863, "dur": 0.828,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034343.000, "dur": 1.618,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034343.507, "dur": 0.849,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034344.863, "dur": 3.594,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034347.456, "dur": 0.933,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034348.742, "dur": 2.765,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034350.443, "dur": 0.746,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034351.866, "dur": 1.072,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034352.221, "dur": 0.648,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034353.189, "dur": 3.178,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034355.740, "dur": 0.560,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034356.592, "dur": 1.933,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034357.855, "dur": 0.603,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034358.785, "dur": 3.585,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034359.876, "dur": 2.424,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034364.746, "dur": 1.870,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034365.904, "dur": 0.643,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034366.838, "dur": 2.213,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034368.000, "dur": 0.984,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034369.298, "dur": 4.029,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034372.338, "dur": 0.922,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034373.575, "dur": 2.202,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034374.836, "dur": 0.877,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034376.016, "dur": 1.648,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034376.833, "dur": 0.764,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034377.888, "dur": 3.955,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034381.117, "dur": 0.660,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034382.107, "dur": 1.851,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034383.150, "dur": 0.643,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034384.179, "dur": 3.703,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034385.611, "dur": 2.202,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034390.643, "dur": 2.050,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034391.778, "dur": 0.848,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034392.915, "dur": 1.935,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034393.930, "dur": 0.850,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034395.130, "dur": 3.690,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034397.996, "dur": 0.753,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034399.068, "dur": 2.008,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034400.252, "dur": 0.752,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034401.339, "dur": 1.528,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034402.002, "dur": 0.797,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034403.089, "dur": 4.220,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034406.511, "dur": 0.732,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034407.547, "dur": 1.791,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034408.516, "dur": 0.752,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034409.562, "dur": 3.977,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034411.002, "dur": 2.467,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034415.866, "dur": 1.611,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034416.737, "dur": 0.677,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034417.702, "dur": 2.143,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034418.963, "dur": 0.815,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034420.207, "dur": 3.560,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034422.971, "dur": 0.731,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034424.128, "dur": 1.672,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034425.031, "dur": 0.701,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034426.067, "dur": 1.548,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034426.900, "dur": 0.648,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034427.835, "dur": 4.319,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034431.108, "dur": 0.787,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034432.396, "dur": 1.739,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034433.154, "dur": 0.914,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034434.361, "dur": 3.972,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034435.821, "dur": 2.439,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034440.507, "dur": 1.508,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034441.187, "dur": 0.760,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034442.238, "dur": 2.357,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034443.436, "dur": 0.868,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034444.854, "dur": 3.633,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034447.734, "dur": 0.685,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034448.744, "dur": 2.029,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034449.864, "dur": 0.658,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034451.037, "dur": 2.074,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034452.192, "dur": 0.650,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034453.333, "dur": 4.027,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034456.551, "dur": 0.740,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034457.624, "dur": 1.717,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034458.540, "dur": 0.735,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034459.663, "dur": 3.508,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034461.008, "dur": 2.089,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034465.680, "dur": 1.862,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034466.607, "dur": 0.666,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034467.793, "dur": 1.901,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034468.908, "dur": 0.724,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034469.960, "dur": 4.431,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034473.709, "dur": 0.611,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034474.675, "dur": 1.764,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034475.599, "dur": 0.771,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034476.693, "dur": 2.004,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034477.751, "dur": 0.881,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034478.949, "dur": 4.543,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034482.572, "dur": 0.853,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034483.736, "dur": 1.888,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034484.719, "dur": 0.842,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034485.866, "dur": 3.448,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034486.930, "dur": 2.316,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034491.530, "dur": 1.687,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034492.390, "dur": 0.761,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034493.457, "dur": 2.032,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034494.745, "dur": 0.679,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034495.771, "dur": 3.547,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034498.464, "dur": 0.787,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034499.565, "dur": 2.006,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034500.813, "dur": 0.689,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034501.808, "dur": 1.688,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034502.568, "dur": 0.863,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034569.591, "dur": 5.802,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034574.361, "dur": 0.645,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034575.812, "dur": 1.330,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034576.544, "dur": 0.528,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034577.473, "dur": 3.645,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034578.584, "dur": 2.460,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034583.658, "dur": 1.725,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034584.549, "dur": 0.765,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034585.747, "dur": 2.546,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034587.056, "dur": 0.970,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034588.630, "dur": 3.509,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034591.504, "dur": 0.564,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034592.449, "dur": 2.016,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034593.827, "dur": 0.568,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034594.741, "dur": 1.639,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034595.394, "dur": 0.917,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034596.668, "dur": 3.911,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034599.965, "dur": 0.544,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034600.985, "dur": 1.681,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034601.776, "dur": 0.619,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034602.890, "dur": 4.042,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034604.407, "dur": 2.261,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034609.638, "dur": 1.612,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034610.491, "dur": 0.691,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034611.548, "dur": 2.060,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034612.886, "dur": 0.654,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034613.855, "dur": 3.659,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034616.736, "dur": 0.709,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034617.779, "dur": 1.732,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034618.639, "dur": 0.803,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034619.752, "dur": 1.647,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034620.641, "dur": 0.687,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034621.730, "dur": 3.803,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034624.648, "dur": 0.813,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034625.785, "dur": 1.688,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034626.629, "dur": 0.775,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034627.726, "dur": 3.438,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034629.105, "dur": 1.980,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034633.525, "dur": 1.429,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034634.259, "dur": 0.621,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034635.185, "dur": 2.253,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034636.618, "dur": 0.753,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034637.751, "dur": 3.727,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034640.673, "dur": 0.734,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034641.987, "dur": 2.219,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034643.377, "dur": 0.564,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034644.471, "dur": 2.180,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034645.398, "dur": 1.007,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034646.893, "dur": 4.204,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034650.326, "dur": 0.701,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034651.352, "dur": 1.256,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034651.977, "dur": 0.562,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034652.857, "dur": 3.478,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034654.047, "dur": 2.218,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034658.510, "dur": 1.340,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034659.137, "dur": 0.642,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034660.091, "dur": 1.927,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034661.227, "dur": 0.724,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034662.264, "dur": 3.199,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034664.824, "dur": 0.569,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034665.753, "dur": 1.963,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034666.916, "dur": 0.733,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034667.958, "dur": 1.598,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034668.614, "dur": 0.872,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034669.809, "dur": 4.222,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034673.194, "dur": 0.770,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034674.376, "dur": 1.949,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034675.282, "dur": 0.651,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034676.563, "dur": 3.168,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034677.568, "dur": 2.088,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034682.366, "dur": 1.657,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034683.187, "dur": 0.768,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034684.562, "dur": 1.709,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034685.509, "dur": 0.695,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034686.545, "dur": 3.450,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034689.433, "dur": 0.491,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034690.244, "dur": 1.803,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034691.297, "dur": 0.679,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034692.287, "dur": 1.754,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034693.069, "dur": 0.904,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034694.335, "dur": 3.692,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034697.455, "dur": 0.506,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034698.298, "dur": 1.513,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034698.933, "dur": 0.812,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034700.045, "dur": 3.494,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034701.227, "dur": 2.238,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034706.003, "dur": 1.602,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034706.938, "dur": 0.599,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034708.232, "dur": 2.110,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034709.401, "dur": 0.873,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034710.790, "dur": 3.627,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034713.734, "dur": 0.611,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034714.667, "dur": 1.880,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034715.895, "dur": 0.585,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034716.862, "dur": 1.516,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034717.601, "dur": 0.709,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034718.618, "dur": 3.629,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034721.577, "dur": 0.589,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034722.598, "dur": 1.738,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034723.447, "dur": 0.822,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034724.773, "dur": 3.923,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034726.185, "dur": 2.439,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034731.135, "dur": 2.024,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034732.542, "dur": 0.539,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034733.638, "dur": 2.120,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034734.891, "dur": 0.795,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034736.024, "dur": 3.278,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034738.549, "dur": 0.686,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034739.568, "dur": 2.003,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034740.902, "dur": 0.600,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034742.009, "dur": 1.523,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034742.797, "dur": 0.670,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034744.022, "dur": 3.893,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034747.124, "dur": 0.726,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034748.386, "dur": 1.313,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034749.049, "dur": 0.582,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034749.987, "dur": 3.704,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034751.195, "dur": 2.234,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034756.034, "dur": 1.552,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034756.738, "dur": 0.780,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034757.838, "dur": 1.820,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034758.978, "dur": 0.610,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034759.925, "dur": 3.325,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034762.657, "dur": 0.521,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034763.494, "dur": 2.112,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034764.833, "dur": 0.707,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034765.845, "dur": 1.768,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034766.820, "dur": 0.725,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034767.862, "dur": 4.375,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034771.321, "dur": 0.848,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034772.512, "dur": 3.392,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034775.034, "dur": 0.803,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034776.167, "dur": 3.524,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034777.231, "dur": 2.376,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034782.344, "dur": 1.782,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034783.277, "dur": 0.780,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034784.349, "dur": 1.971,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034785.475, "dur": 0.775,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034786.610, "dur": 3.418,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034789.284, "dur": 0.678,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034790.269, "dur": 1.772,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034791.345, "dur": 0.629,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034792.294, "dur": 1.609,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034793.347, "dur": 0.487,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034794.163, "dur": 4.506,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034797.800, "dur": 0.800,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034798.943, "dur": 1.333,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034799.733, "dur": 0.473,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034800.497, "dur": 4.080,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034801.883, "dur": 2.620,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034806.972, "dur": 1.412,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034807.583, "dur": 0.733,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034808.808, "dur": 2.394,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034810.282, "dur": 0.851,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034811.443, "dur": 3.351,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034814.062, "dur": 0.662,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034815.036, "dur": 2.202,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034816.429, "dur": 0.739,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034817.478, "dur": 1.612,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034818.181, "dur": 0.841,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034819.533, "dur": 4.367,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034822.962, "dur": 0.866,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034824.142, "dur": 1.707,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034825.024, "dur": 0.758,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034826.271, "dur": 3.602,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034827.345, "dur": 2.458,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034832.298, "dur": 1.560,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034833.271, "dur": 0.517,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034834.436, "dur": 2.257,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034835.725, "dur": 0.897,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034836.945, "dur": 3.456,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034839.719, "dur": 0.612,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034840.654, "dur": 1.644,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034841.628, "dur": 0.607,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034842.538, "dur": 1.430,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034843.197, "dur": 0.704,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034844.190, "dur": 3.532,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034847.088, "dur": 0.566,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034848.079, "dur": 1.581,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034848.793, "dur": 0.801,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034849.881, "dur": 3.899,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034851.488, "dur": 2.124,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034855.886, "dur": 1.754,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034857.024, "dur": 0.550,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034858.228, "dur": 1.749,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034859.228, "dur": 0.681,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034860.386, "dur": 3.314,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034863.205, "dur": 0.423,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034863.951, "dur": 2.060,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034864.999, "dur": 0.941,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034866.289, "dur": 1.370,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034867.013, "dur": 0.575,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034868.065, "dur": 3.415,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034870.582, "dur": 0.832,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034872.059, "dur": 1.274,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034872.697, "dur": 0.566,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034873.553, "dur": 3.385,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034874.830, "dur": 2.020,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034879.388, "dur": 1.509,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034880.061, "dur": 0.766,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034881.182, "dur": 2.336,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034882.595, "dur": 0.857,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034883.754, "dur": 3.270,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034886.504, "dur": 0.451,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034887.264, "dur": 2.044,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034888.592, "dur": 0.653,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034889.737, "dur": 1.529,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034890.486, "dur": 0.712,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034891.488, "dur": 4.745,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034895.193, "dur": 0.972,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034896.488, "dur": 1.511,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034897.194, "dur": 0.738,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034898.224, "dur": 3.486,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034899.557, "dur": 2.079,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034904.018, "dur": 1.393,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034904.741, "dur": 0.604,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034905.653, "dur": 1.675,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034906.585, "dur": 0.675,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034907.781, "dur": 3.737,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034910.698, "dur": 0.751,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034911.817, "dur": 1.865,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034912.965, "dur": 0.652,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034913.921, "dur": 1.553,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034914.730, "dur": 0.674,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034915.695, "dur": 3.536,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034918.554, "dur": 0.606,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034919.507, "dur": 1.626,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034920.434, "dur": 0.630,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034921.566, "dur": 3.710,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034923.190, "dur": 1.822,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034927.487, "dur": 1.781,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034928.437, "dur": 0.764,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034929.539, "dur": 2.378,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034930.984, "dur": 0.868,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034932.302, "dur": 4.004,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034935.478, "dur": 0.753,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034936.552, "dur": 2.031,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034937.925, "dur": 0.585,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034938.876, "dur": 1.837,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034939.792, "dur": 0.852,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034940.934, "dur": 4.795,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034945.076, "dur": 0.584,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034945.967, "dur": 1.814,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034946.821, "dur": 0.893,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034948.209, "dur": 3.281,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034949.265, "dur": 2.150,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034954.803, "dur": 1.419,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034955.416, "dur": 0.736,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034956.518, "dur": 2.384,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034957.772, "dur": 1.064,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034959.150, "dur": 3.632,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034962.107, "dur": 0.606,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034963.062, "dur": 2.018,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034964.448, "dur": 0.560,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034965.325, "dur": 1.554,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034966.185, "dur": 0.623,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034967.098, "dur": 4.122,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034970.394, "dur": 0.755,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034971.456, "dur": 1.330,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034972.147, "dur": 0.572,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034973.012, "dur": 4.066,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034974.503, "dur": 2.504,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034979.429, "dur": 1.358,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034980.094, "dur": 0.625,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658034981.225, "dur": 26.834,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035006.277, "dur": 1.332,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035008.856, "dur": 3.726,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035011.897, "dur": 0.615,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035012.859, "dur": 2.287,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035014.507, "dur": 0.572,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035015.410, "dur": 1.941,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035016.323, "dur": 0.955,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035017.619, "dur": 4.369,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035021.205, "dur": 0.715,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035022.275, "dur": 1.624,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035023.137, "dur": 0.692,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035024.205, "dur": 3.640,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035025.355, "dur": 2.413,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035030.818, "dur": 1.620,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035031.624, "dur": 0.743,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035032.725, "dur": 2.059,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035033.818, "dur": 0.892,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035035.083, "dur": 3.621,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035038.078, "dur": 0.558,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035038.952, "dur": 1.828,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035039.856, "dur": 0.860,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035041.245, "dur": 1.567,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035042.072, "dur": 0.668,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035043.046, "dur": 4.121,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035046.344, "dur": 0.750,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035047.587, "dur": 1.702,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035048.395, "dur": 0.824,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035049.520, "dur": 3.787,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035051.047, "dur": 2.182,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035055.560, "dur": 1.371,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035056.249, "dur": 0.615,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035057.393, "dur": 1.931,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035058.563, "dur": 0.687,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035059.703, "dur": 3.673,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035062.579, "dur": 0.727,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035063.621, "dur": 2.213,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035065.018, "dur": 0.753,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035066.091, "dur": 1.623,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035066.775, "dur": 0.868,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035068.318, "dur": 3.831,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035071.448, "dur": 0.635,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035072.450, "dur": 1.579,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035073.198, "dur": 0.758,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035074.257, "dur": 3.501,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035075.460, "dur": 2.219,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035080.699, "dur": 1.664,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035081.523, "dur": 0.773,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035082.782, "dur": 2.403,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035084.215, "dur": 0.898,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035085.448, "dur": 3.663,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035088.490, "dur": 0.554,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035089.382, "dur": 1.732,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035090.460, "dur": 0.581,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035091.365, "dur": 1.582,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035092.116, "dur": 0.762,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035093.199, "dur": 3.723,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035096.171, "dur": 0.681,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035097.475, "dur": 1.789,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035098.413, "dur": 0.781,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035099.510, "dur": 3.819,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035100.832, "dur": 2.423,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035105.938, "dur": 1.785,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035106.861, "dur": 0.794,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035108.147, "dur": 2.328,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035109.485, "dur": 0.919,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035110.868, "dur": 3.620,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035113.778, "dur": 0.637,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035114.768, "dur": 1.952,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035116.045, "dur": 0.612,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035116.971, "dur": 2.085,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035117.771, "dur": 1.213,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035119.288, "dur": 3.147,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035121.748, "dur": 0.620,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035122.692, "dur": 1.904,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035123.754, "dur": 0.774,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035125.097, "dur": 3.603,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035126.260, "dur": 2.273,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035131.056, "dur": 1.394,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035131.694, "dur": 0.687,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035132.713, "dur": 2.389,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035133.919, "dur": 1.108,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035135.491, "dur": 3.329,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035138.267, "dur": 0.486,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035139.063, "dur": 2.215,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035140.428, "dur": 0.784,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035141.529, "dur": 1.861,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035142.424, "dur": 0.892,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035143.795, "dur": 3.123,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035146.392, "dur": 0.454,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035147.222, "dur": 1.672,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035148.146, "dur": 0.677,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035149.124, "dur": 3.630,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035150.508, "dur": 2.171,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035154.988, "dur": 1.481,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035155.656, "dur": 0.743,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035156.715, "dur": 2.037,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035157.893, "dur": 0.791,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035159.255, "dur": 3.487,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035161.956, "dur": 0.716,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035162.991, "dur": 1.594,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035163.882, "dur": 0.637,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035164.868, "dur": 1.876,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035165.948, "dur": 0.724,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035167.168, "dur": 3.969,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035170.257, "dur": 0.810,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035171.388, "dur": 1.543,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035172.003, "dur": 0.857,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035173.408, "dur": 3.938,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035174.738, "dur": 2.530,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035179.633, "dur": 1.311,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035180.292, "dur": 0.581,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035181.541, "dur": 1.662,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035182.432, "dur": 0.697,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035183.471, "dur": 3.513,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035186.257, "dur": 0.658,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035187.241, "dur": 2.003,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035188.699, "dur": 0.474,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035189.496, "dur": 1.393,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035190.122, "dur": 0.693,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035191.116, "dur": 4.111,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035194.526, "dur": 0.638,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035195.826, "dur": 2.761,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035197.810, "dur": 0.708,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035221.822, "dur": 136.004,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035459.571, "dur": 158.703,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035525.688, "dur": 69.534,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658035555.655, "dur": 1.241,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658036031.185, "dur": 1022.025,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658036090.427, "dur": 133.804,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658036096.039, "dur": 1.374,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658036167.635, "dur": 0.894,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658036253.319, "dur": 442.171,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658036260.132, "dur": 45.224,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658036263.419, "dur": 8.616,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658036267.248, "dur": 3.881,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658036273.712, "dur": 31.043,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658036315.321, "dur": 375.910,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658036346.627, "dur": 337.927,
    "args": {
      "External id": 2489747,"Record function id": 0, "Collective name": "allreduce", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 0, "Input Strides": [[[]], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "0", "", "[]", "[]", "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": 1336753, "tid": 1336753,
    "ts": 1514658036364.160, "dur": 314.123,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658036766.327, "dur": 210.429,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658036861.484, "dur": 32.129,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658036881.104, "dur": 3.946,
    "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": ["", "", "0", "", "[]", "[]", "-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": 0, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 1336753, "tid": 1336753,
    "ts": 1514658036924.687, "dur": 45.708,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658036927.521, "dur": 1.083,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658036929.839, "dur": 0.527,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658037071.878, "dur": 33.502,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658037082.712, "dur": 16.466,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658037085.096, "dur": 13.044,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658037470.643, "dur": 41.417,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658037521.515, "dur": 43.102,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658037573.932, "dur": 36.913,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658037628.232, "dur": 28.712,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658037631.887, "dur": 0.519,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658037676.856, "dur": 0.688,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658037792.449, "dur": 751.810,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658038328.737, "dur": 169.096,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658038594.468, "dur": 36.478,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658038599.103, "dur": 31.354,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658038634.802, "dur": 478.355,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658038636.659, "dur": 476.278,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658038638.494, "dur": 473.589,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658039119.834, "dur": 65.497,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658039123.218, "dur": 37.933,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658039131.444, "dur": 3.440,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658039136.532, "dur": 24.091,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658039141.904, "dur": 2.814,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658039162.947, "dur": 21.422,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658039187.812, "dur": 37.894,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658039191.176, "dur": 34.385,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658039192.233, "dur": 33.057,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658039264.075, "dur": 6142.621,
    "args": {
      "External id": 2489780,"Record function id": 0, "Ev Idx": 11339
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "Optimizer.step#AdamW.step", "pid": 1336753, "tid": 1336753,
    "ts": 1514658039297.002, "dur": 6086.272,
    "args": {
      "External id": 2489781,"Record function id": 0, "Ev Idx": 11340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_add_", "pid": 1336753, "tid": 1336753,
    "ts": 1514658040686.588, "dur": 384.397,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040706.566, "dur": 0.950,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040708.422, "dur": 0.118,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040708.973, "dur": 0.076,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040709.680, "dur": 0.071,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040710.120, "dur": 0.267,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040710.782, "dur": 0.068,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040711.193, "dur": 0.314,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040711.837, "dur": 0.274,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040712.667, "dur": 0.059,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040713.104, "dur": 0.246,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040713.984, "dur": 0.044,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040714.393, "dur": 0.250,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040715.000, "dur": 0.059,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040715.449, "dur": 0.059,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040715.877, "dur": 0.057,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040716.297, "dur": 0.060,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040716.740, "dur": 0.055,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040717.179, "dur": 0.099,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040717.615, "dur": 0.059,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040718.065, "dur": 0.048,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040718.512, "dur": 0.048,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040718.888, "dur": 0.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040719.306, "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": 1336753, "tid": 1336753,
    "ts": 1514658040719.729, "dur": 0.062,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040720.336, "dur": 0.061,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040720.733, "dur": 0.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040721.206, "dur": 0.052,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040721.633, "dur": 0.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040722.093, "dur": 0.094,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040722.540, "dur": 0.052,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040723.164, "dur": 0.244,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040723.887, "dur": 0.252,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040724.579, "dur": 0.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040725.066, "dur": 0.287,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040725.704, "dur": 0.253,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040726.484, "dur": 0.283,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040727.232, "dur": 0.282,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040727.960, "dur": 0.066,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040728.367, "dur": 0.066,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040728.806, "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": 1336753, "tid": 1336753,
    "ts": 1514658040729.283, "dur": 0.059,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040729.737, "dur": 0.273,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040730.511, "dur": 0.055,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040731.049, "dur": 0.057,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040731.444, "dur": 0.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040732.023, "dur": 0.082,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040732.662, "dur": 0.053,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040733.054, "dur": 0.059,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040733.441, "dur": 0.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040733.804, "dur": 0.055,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040734.245, "dur": 0.059,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040734.969, "dur": 0.062,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040735.361, "dur": 0.061,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040735.765, "dur": 0.310,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040736.364, "dur": 0.276,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040736.984, "dur": 0.230,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040737.630, "dur": 0.070,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040738.077, "dur": 0.269,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040738.762, "dur": 0.261,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040739.404, "dur": 0.082,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040739.831, "dur": 0.302,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040740.511, "dur": 0.274,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040741.175, "dur": 0.174,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040741.729, "dur": 0.119,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040742.128, "dur": 0.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040742.486, "dur": 0.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040742.913, "dur": 0.059,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040743.308, "dur": 0.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040743.787, "dur": 0.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040744.334, "dur": 0.057,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040744.732, "dur": 0.057,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040745.212, "dur": 0.054,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040745.692, "dur": 0.051,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040746.103, "dur": 0.059,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040820.828, "dur": 0.397,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040822.641, "dur": 0.067,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040823.076, "dur": 0.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040823.592, "dur": 0.049,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040823.980, "dur": 0.061,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040824.439, "dur": 0.060,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040824.848, "dur": 0.060,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040825.294, "dur": 0.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040825.680, "dur": 0.057,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040826.145, "dur": 0.074,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040826.588, "dur": 0.053,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040827.014, "dur": 0.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040827.427, "dur": 0.069,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040827.883, "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": 1336753, "tid": 1336753,
    "ts": 1514658040828.303, "dur": 0.056,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040828.703, "dur": 0.057,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040829.274, "dur": 0.047,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040829.716, "dur": 0.060,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040830.188, "dur": 0.071,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040830.607, "dur": 0.056,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040831.059, "dur": 0.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040831.372, "dur": 0.054,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040831.766, "dur": 0.057,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040832.311, "dur": 0.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040832.719, "dur": 0.053,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040833.141, "dur": 0.055,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040833.548, "dur": 0.053,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040833.957, "dur": 0.056,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040834.392, "dur": 0.056,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040834.821, "dur": 0.059,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040835.209, "dur": 0.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040835.599, "dur": 0.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040836.079, "dur": 0.045,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040836.435, "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": 1336753, "tid": 1336753,
    "ts": 1514658040836.836, "dur": 0.056,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040837.325, "dur": 0.060,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040837.788, "dur": 0.063,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040838.139, "dur": 0.057,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040838.547, "dur": 0.064,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040838.933, "dur": 0.061,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040839.362, "dur": 0.056,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040839.776, "dur": 0.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040840.221, "dur": 0.059,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040840.558, "dur": 0.057,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040840.947, "dur": 0.060,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040841.347, "dur": 0.060,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040841.740, "dur": 0.060,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040842.098, "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": 1336753, "tid": 1336753,
    "ts": 1514658040842.515, "dur": 0.061,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040842.913, "dur": 0.059,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040843.321, "dur": 0.055,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040843.747, "dur": 0.057,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040844.178, "dur": 0.055,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040868.274, "dur": 0.238,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040869.664, "dur": 0.082,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040870.110, "dur": 0.064,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040870.559, "dur": 0.054,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040870.992, "dur": 0.061,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040871.350, "dur": 0.053,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040871.712, "dur": 0.056,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040872.092, "dur": 0.055,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040872.559, "dur": 0.054,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040873.056, "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": 1336753, "tid": 1336753,
    "ts": 1514658040873.470, "dur": 0.064,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040873.876, "dur": 0.059,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040874.351, "dur": 0.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040874.758, "dur": 0.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040875.165, "dur": 0.056,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040875.771, "dur": 0.055,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040876.125, "dur": 0.056,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040876.517, "dur": 0.055,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040876.952, "dur": 0.060,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040877.360, "dur": 0.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040877.685, "dur": 0.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040878.148, "dur": 0.055,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040878.606, "dur": 0.060,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040879.040, "dur": 0.050,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040879.408, "dur": 0.057,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040879.730, "dur": 0.059,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040880.045, "dur": 0.053,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040880.464, "dur": 0.061,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040880.869, "dur": 0.060,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040881.298, "dur": 0.057,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040881.813, "dur": 0.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040882.168, "dur": 0.053,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040882.558, "dur": 0.059,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040882.959, "dur": 0.055,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040883.381, "dur": 0.066,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040883.776, "dur": 0.068,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040884.228, "dur": 0.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040884.599, "dur": 0.059,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040884.988, "dur": 0.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040885.363, "dur": 0.059,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040885.779, "dur": 0.055,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040886.184, "dur": 0.071,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040886.628, "dur": 0.057,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040887.067, "dur": 0.067,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040887.521, "dur": 0.062,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040887.882, "dur": 0.069,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040888.340, "dur": 0.065,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040888.774, "dur": 0.068,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040889.185, "dur": 0.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040889.782, "dur": 0.057,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040890.241, "dur": 0.055,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040890.632, "dur": 0.061,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040891.118, "dur": 0.060,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040891.731, "dur": 0.059,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040892.156, "dur": 0.062,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040892.571, "dur": 0.054,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040893.030, "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": 1336753, "tid": 1336753,
    "ts": 1514658040893.555, "dur": 0.059,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040893.958, "dur": 0.055,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040894.482, "dur": 0.056,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040894.834, "dur": 0.057,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040895.228, "dur": 0.057,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040895.638, "dur": 0.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040896.100, "dur": 0.057,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040896.535, "dur": 0.055,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040896.890, "dur": 0.055,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040897.370, "dur": 0.047,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040897.819, "dur": 0.059,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040898.623, "dur": 0.060,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040899.028, "dur": 0.057,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040899.351, "dur": 0.056,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040899.765, "dur": 0.057,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040900.128, "dur": 0.057,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040900.704, "dur": 0.057,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040901.087, "dur": 0.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040901.470, "dur": 0.066,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040901.843, "dur": 0.063,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040902.256, "dur": 0.058,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040902.670, "dur": 0.069,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040903.034, "dur": 0.062,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040903.430, "dur": 0.069,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040903.776, "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": 1336753, "tid": 1336753,
    "ts": 1514658040904.144, "dur": 0.067,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040904.540, "dur": 0.086,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040905.201, "dur": 0.060,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040905.710, "dur": 0.061,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040906.150, "dur": 0.064,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040906.538, "dur": 0.067,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040906.998, "dur": 0.063,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040907.524, "dur": 0.060,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040907.950, "dur": 0.054,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040908.364, "dur": 0.064,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040908.794, "dur": 0.063,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040909.278, "dur": 0.066,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040909.721, "dur": 0.062,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040910.214, "dur": 0.065,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040910.607, "dur": 0.060,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040910.991, "dur": 0.060,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040911.427, "dur": 0.052,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040912.374, "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": 1336753, "tid": 1336753,
    "ts": 1514658040912.870, "dur": 0.062,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040913.285, "dur": 0.060,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040913.690, "dur": 0.059,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040914.081, "dur": 0.056,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040914.401, "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": 1336753, "tid": 1336753,
    "ts": 1514658040914.906, "dur": 0.057,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040915.250, "dur": 0.057,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040915.566, "dur": 0.052,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040916.245, "dur": 0.056,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040916.636, "dur": 0.054,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040917.237, "dur": 0.062,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040917.537, "dur": 0.050,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040918.331, "dur": 0.072,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040918.645, "dur": 0.052,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040919.377, "dur": 0.066,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040919.690, "dur": 0.075,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040920.325, "dur": 0.071,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040920.641, "dur": 0.050,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040921.366, "dur": 0.059,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040921.670, "dur": 0.049,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040922.431, "dur": 0.059,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040922.734, "dur": 0.053,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040923.588, "dur": 0.065,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040923.897, "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": 1336753, "tid": 1336753,
    "ts": 1514658040924.856, "dur": 0.060,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040925.159, "dur": 0.053,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040925.711, "dur": 0.065,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040926.025, "dur": 0.051,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040926.594, "dur": 0.067,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040926.903, "dur": 0.055,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040927.638, "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": 1336753, "tid": 1336753,
    "ts": 1514658040927.968, "dur": 0.053,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040928.497, "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": 1336753, "tid": 1336753,
    "ts": 1514658040928.809, "dur": 0.051,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040929.638, "dur": 0.064,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040929.949, "dur": 0.049,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040930.968, "dur": 0.059,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040931.272, "dur": 0.053,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040932.089, "dur": 0.063,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040932.414, "dur": 0.048,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040933.229, "dur": 0.067,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040933.542, "dur": 0.054,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040934.357, "dur": 0.065,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040934.669, "dur": 0.056,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040935.516, "dur": 0.068,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040935.830, "dur": 0.050,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040936.945, "dur": 0.056,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040937.253, "dur": 0.055,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040938.094, "dur": 0.091,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040938.435, "dur": 0.049,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040939.027, "dur": 0.067,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040939.340, "dur": 0.043,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040940.045, "dur": 0.057,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040940.350, "dur": 0.049,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040941.195, "dur": 0.054,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040941.494, "dur": 0.046,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040942.285, "dur": 0.057,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040942.638, "dur": 0.046,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040943.595, "dur": 0.056,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040943.922, "dur": 0.044,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040944.731, "dur": 0.057,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040945.030, "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": 1336753, "tid": 1336753,
    "ts": 1514658040945.853, "dur": 0.065,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658040946.262, "dur": 0.053,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658041602.347, "dur": 3689.692,
    "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": 1336753, "tid": 1336753,
    "ts": 1514658044689.182, "dur": 405.086,
    "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": 1514655937455.088, "pid": 1336753, "tid": 0,
    "args": {
      "name": "python3.12"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1514655937455.088, "pid": 1336753, "tid": 0,
    "args": {
      "labels": "CPU"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1514655937455.088, "pid": 1336753, "tid": 0,
    "args": {
      "sort_index": 1336753
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 1514655937455.088, "pid": 1336753, "tid": 1381210,
    "args": {
      "name": "thread 1381210 (python3.12)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 1514655937455.088, "pid": 1336753, "tid": 1381210,
    "args": {
      "sort_index": 1381210
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 1514655937455.088, "pid": 1336753, "tid": 1336753,
    "args": {
      "name": "thread 1336753 (python3.12)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 1514655937455.088, "pid": 1336753, "tid": 1336753,
    "args": {
      "sort_index": 1336753
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 1514655937455.088, "pid": 1336753, "tid": 1381210,
    "args": {
      "name": "thread 1381210 (pt_autograd_0)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 1514655937455.088, "pid": 1336753, "tid": 1381210,
    "args": {
      "sort_index": 1381210
    }
  },
  {
    "ph": "X", "cat": "Trace", "ts": 1514655937388.489, "dur": 2109573.906,
    "pid": "Spans", "tid": "PyTorch Profiler",
    "name": "PyTorch Profiler (0)",
    "args": {
      "Op count": 0
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1514655937388.489,
    "pid": "Spans", "tid": 0,
    "args": {
      "sort_index": 536870912
    }
  },
  {
    "name": "Iteration Start: PyTorch Profiler", "ph": "i", "s": "g",
    "pid": "Traces", "tid": "Trace PyTorch Profiler", "ts": 1514655937388.489
  },
  {
    "name": "Record Window End", "ph": "i", "s": "g",
    "pid": "", "tid": "", "ts": 1514658122447.425
  }
  ],
  "traceName": "exp/mtp.1B.batch16.seqlen4096.context4096.warmup2000.update1.steps200000.lr2e-4.cosine/profile_trace/iteration_102912/rank0_trace.json",
  "displayTimeUnit": "ms",
  "baseTimeNanoseconds": 1751410836000000000
}